I am an engineering student working for the first time in breast cancer field. I am trying to learn breast cancer scoring (for example Allred score). I have come across 2 types of images:
First (let be A):
Second (let be B):
Images are from the VitroVivo site.
My confusion: For breast cancer scoring, is image A used?
What I think I understand: In cancer scoring (in breast), we are trying to find a score. For example in image B we have immunopositive (brown) nuclei and immunonegative (blue) nuclei and the ratio of brown to total is an indicator. This brown nuclei can be Ki67, HER2, Estrogen, Progesterone protein/antigen. But in H&E stained images, nuclei are separated from cytoplasm. There are no positive/negative nuclei, only one type of nucleus. So, is there any way H&E stained tissues are used for scoring?
Immunohistochemistry and Immunocytochemistry
Counterstains for Enzyme/Chromogen Immunostaining
Hematoxylin is arguably the most common nuclear counterstain used when employing an enzyme/chromogen detection system. There are various formulations available, classified by the type of mordant used and whether they are progressive or regressive. All of them ultimately give cell nuclei a pleasing blue coloration of varying hue and intensity, depending on the type of hematoxylin used.
Hematoxylin alone (or more accurately its oxidation product, hematin) is anionic and therefore does not have much affinity for DNA. Mordants are iron salts, namely those of iron, aluminum, tungsten, and lead. Mordants combine with hematin, resulting in a positively charged dye–mordant complex, thus allowing it to bind to anionic chromatin. Alum (aluminum mordanted) hematoxylins can be used progressively or regressively. With progressive hematoxylins (such as Mayer’s, Carazzi’s, and Gill’s), tissue or cells are incubated in hematoxylin until the desired degree of nuclear staining is achieved, before being blued. In comparison, in the case of regressive hematoxylins (such as Harris’s), tissues or cells are incubated until a degree of overstaining is achieved, before having some of the excess hematoxylin removed by immersion in an acid solution, such as 1% acid alcohol. This process is known as differentiation. Progressive hematoxylins are therefore more convenient to use than regressive, due to the absence of a differentiation step and the resulting compatibility with alcohol-soluble enzyme/substrate end products, such as those produced by HRP and AEC.
Whether progressive or regressive, once the desired level of nuclear staining is achieved, hematoxylins are “blued.” At acid pH, hematoxylins stain the nuclei red. However, once exposed to an alkaline environment, hematoxylin turns a pleasing blue color. Running tap water is commonly used for this purpose since it has sufficient alkalinity, especially in “hard” water areas. In areas of “soft” water, a suitable alkaline solution can be used to blue hematoxylin, such as 0.05% (v/v) ammonia.
Other commonly used tinctorial nuclear counterstains are light green, fast red, toluidine blue, and methylene blue staining nuclei either green, red, or blue, respectively.
One important consideration when using a nuclear tinctorial counterstain is not to make the staining too intense if you are demonstrating a nuclear antigen, since the counterstain can potentially mask the positive signal from the detection system.
Hematoxylin and Eosin Staining of Tissue and Cell Sections
Hematoxylin and eosin (H&E) stains have been used for at least a century and are still essential for recognizing various tissue types and the morphologic changes that form the basis of contemporary cancer diagnosis. The stain has been unchanged for many years because it works well with a variety of fixatives and displays a broad range of cytoplasmic, nuclear, and extracellular matrix features. Hematoxylin has a deep blue-purple color and stains nucleic acids by a complex, incompletely understood reaction. Eosin is pink and stains proteins nonspecifically. In a typical tissue, nuclei are stained blue, whereas the cytoplasm and extracellular matrix have varying degrees of pink staining. Well-fixed cells show considerable intranuclear detail. Nuclei show varying cell-type- and cancer-type-specific patterns of condensation of heterochromatin (hematoxylin staining) that are diagnostically very important. Nucleoli stain with eosin. If abundant polyribosomes are present, the cytoplasm will have a distinct blue cast. The Golgi zone can be tentatively identified by the absence of staining in a region next to the nucleus. Thus, the stain discloses abundant structural information, with specific functional implications. A limitation of hematoxylin staining is that it is incompatible with immunofluorescence. It is useful, however, to stain one serial paraffin section from a tissue in which immunofluorescence will be performed. Hematoxylin, generally without eosin, is useful as a counterstain for many immunohistochemical or hybridization procedures that use colorimetric substrates (such as alkaline phosphatase or peroxidase). This protocol describes H&E staining of tissue and cell sections.
Staining of tissues sections using chemical and biological dyes has been used for over a century for visualizing various tissue types and morphologic changes associated with contemporary cancer diagnosis. The staining procedure however is labor intensive, needs trained technicians, costly, and often results in loss of irreplaceable specimen and delays diagnoses. In collaboration with Brigham and Women's Hospital (Boston, MA), we describe a “computational staining” approach to digitally stain photographs of unstained tissue biopsies with Haematoxylin and Eosin (H&E) dyes to diagnose cancer.
Our method uses neural networks to rapidly stain photographs of non-stained tissues, providing physicians timely information about the anatomy and structure of the tissue. We also report a "computational destaining" algorithm that can remove dyes and stains from photographs of previously stained tissues, allowing reuse of patient samples.
These methods and neural networks assist physicians and patients by novel computational processes at the point-of-care, which can integrate seamlessly into clinical workflows in hospitals all over the world.
Results and Discussion
Effectiveness of iHE
We established a simple but effective staining system (Fig. 1A). The tissues were immersed in dye solution in a centrifuge tube, which was fixed in a stainless-steel container. An ultrasonic transducer was adhered to the underside of a stainless-steel container and a radiator, which was used to cool the ultrasonic transducer. The stainless-steel container was filled with water to ensure better ultrasound transmission. The maximum electric power of the ultrasonic transducer was 60 watts. During the process of staining the tissue with hematoxylin or rinsing, we adjusted the input voltage of the ultrasonic power supply and controlled the acoustic power density in the stainless-steel container at 1.2
1.5 W/cm 2 . During tissue staining with eosin, the acoustic power density was 0.8
Generally, it is difficult to stain intact tissue uniformly (Fig. 1C). Here, we used two means to enhance dyeing—DCM delipidation and ultrasound. DCM is a solvent that can dissolve the lipids in the cell membrane easily 5,11,15,31 . A comparison of Fig. 1C-I with Fig. 1C-III shows that the brain treated with DCM delipidation could be stained more deeply, likely because the brain becomes porous and dye diffusion tends to be enhanced after delipidation. A comparison of Fig. 1C-III with Fig. 1C-IV shows that the brain stained under ultrasound had more uniform staining in a shorter time. In Fig. 1C-II, the image shows that the brain tissue treated with DCM delipidation and stained under ultrasound for 6 h presented uniform hematoxylin staining. In other words, delipidation and ultrasound were indispensable for iHE, and both contributed to the staining results. Comparing Fig. 1C-II with Fig. 1C-III, we also found that encephalocele of the brain with ultrasound and delipidation was larger than that of the brain without ultrasound and delipidation, possibly due to the shrinkage of the brain during dehydration and delipidation.
Another effect of ultrasound was accelerated rinsing (Fig. 1B). Adult mouse brains with delipidation were first stained by hematoxylin for 6 h and then cut at the midline. One half was rinsed under ultrasound, while the other half was rinsed without ultrasound. The rinsing temperature was 50 °C. To obtain the absorbance of the ultrasound rinsing solution and static rinsing solution (control), we transferred the upper solution and measured the absorbance using the Lambda 950 UV/VIS spectrometer at a wavelength of 445 nm. Over time, the absorbance of the ultrasound rinsing solution became much higher than that of the static rinsing solution. Thus, the brain under ultrasound released more dyes than that in the static state.
Comparison of iHE with traditional H&E staining
Ultrasound can lead to tissue and cell disruption thus, a non-negligible issue is whether the cell structure is destroyed after iHE. As shown in Fig. 2, we found that the cell structure was preserved without significant distortion compared with that of traditional H&E staining. The staining effects of iHE and traditional H&E staining similar, indicating that iHE is a feasible method. The perisomatic nerve fibers shown in Fig. 2D are fewer than those in Fig. 2A, likely because the cytomembrane was partially dissolved after delipidation. To further evaluate the sacrifice of subcellular structure, we compared 100 neuron nuclei from the cortex using the H&E and iHE methods, but we found no significant difference between these two methods (data not shown).
Comparison of iHE and traditional H&E staining. (A–C), Images of 7-μm-thick mouse brain slices stained using the traditional H&E method. (D–F), Images of intact mouse brain tissues after iHE staining, slicing and 2D imaging. The red matter appearing in (A
F) represents blood cells that were not completely cleared during cardiac perfusion. Objective lens, 20× N.A., 0.75 work distance, 1 mm. Scale bar: 50 μm.
Intact mouse brain stained with iHE
An intact C57BL/6 mouse brain was stained with iHE, and the images in Fig. 3 show that the brain was stained uniformly. The hippocampus was distinguishable, and the cellular nuclei morphology was clear (Fig. 3D–F). Thus, we could acquire H&E staining information for a mouse brain based on its natural spatial context. Encephalocoele of the brain was apparent (Fig. 3C).
Slices from an intact mouse brain stained with iHE, followed by 2D imaging. (A) 3D projection of 20 slices in (B). (B) The coronal plane slices from the C57BL/6 mouse brain after iHE. Objective lens, 4× and 20× N.A., 0.2 and 0.75. The brain was sliced at the coronal plane for 8 μm, and we selected one slice for every 400 μm from the olfactory bulb to the epencephalon. (D–F) Magnification of (C).
Other intact mouse tissues stained with iHE
To generate a liver metastasis model, we injected 4T-1 mammary cancer cells into 7-week-old BALB/c mice, reared the mice for 40 days and perfused the mice at 3 months of age. Next, the liver was stained, and slices were prepared for subsequent imaging using Nikon NIS-Elements microscopy the results are shown in Fig. 4A–C. In the image, the black dotted line is used to divide the tumor area and normal area, and the results show that the nuclear-cytoplasmic ratio of the tumor area was different from that of the normal area. The mouse lung, kidney, stomach, forepaw, heart and eyeballs were stained using iHE (Figs 4D–G and S4A–H). The pulmonary lobe and alveoli, classic lung structures, can be observed in Fig. 4D,E. Kidney tubules are presented clearly in Fig. 4F,G. The mouse stomach was stained, and the villus at the pylorus is shown in Supplementary Fig. 4G,H. The classical muscle structure in the mouse forepaw was revealed after iHE, and the skin microstructure was also observed (Supplementary Fig. S4A,B). The heart is an organ comprised of cardiomyocytes, and chamber and myocardial cell types were clearly observed (Supplementary Fig. S4C,D). We also applied iHE to mouse eyeballs, and the cellular stratification of the retina could be observed (Supplementary Fig. S4E,F).
Images of other mouse tissues stained with iHE. (A–C) Images of mouse liver with a tumor. (D,E) iHE for mouse lung. (F,G) iHE for mouse kidney. Objective lens, 20× N.A., 0.75 work distance, 1 mm.
Compatibility of iHE with blood vessel staining
Cancer cell growth is associated with significant changes in blood vessels 30 . As discussed above, H&E staining is a classic means to present the details of a tumor. Here, we combined H&E staining with blood vessel staining (Fig. 5) to observe blood vessel and cellular nuclei information simultaneously. This double staining method has the potential to provide meaningful information on tumor status.
Mouse brain perfused with carbon ink before iHE. (A,B) Hippocampus, (B) is the magnification of the box in (A). (C,D) Thalami, (D) is the magnification of the box in (C). Objective lens, 20× N.A., 0.75 work distance, 1 mm.
The principle of iHE is shown in Fig. 6. After DCM delipidation, the cytomembrane is partly dissolved, and the tissue becomes porous 25 . We assume the ultrasound causes stable cavitation (ultrasound energy density <10 Watt/cm 2 ). The pores of the tissue may change their diameter in the ultrasonic field because ultrasound is a longitudinal wave. Similar to a spring stretching and compressing due to an external force, the tissue is stretched and compressed at the microscale level. Thus, the pores in the wave crest of the ultrasound are stretched, while the pores in the trough of the waves are compressed 26,32,33 . Meanwhile, with the ultrasonic wave, there is periodic compression and rarefaction in the liquid. Thus, mass transfer in and out of the tissue is facilitated 34 . Therefore, the diffusion driven by the concentration difference in the particles is accelerated in the ultrasonic field, and the dyes diffuse into tissue faster.
Possible principle of iHE. A fixed tissue is used as the original tissue. After dehydration, the tissue is subjected to delipidation, followed by the H&E staining procedure the H&E staining process was simplified so that the reader could better understand the principle for how ultrasound is involved in the process. Dichloromethane can cause a porous state of the cell membrane, equal to the porous state of the tissue. The pores of the tissue change in the ultrasonic field because ultrasound is a longitudinal wave. Thus, the pores in the wave crest of the ultrasound are stretched, while the pores in the trough of the waves are compressed. Meanwhile, the random motion of particles is enhanced in the ultrasonic field. The diffusion process is enhanced and achieves rapid and uniform staining.
Overall, the iHE method creates porous tissue for staining and enhances random motion. Therefore, the diffusion process is enhanced and achieves rapid and uniform H&E staining. We established a set of methods (iHE) to obtain H&E staining information of intact tissue for volume imaging. Ultrasound was used to facilitate the staining of whole-mount tissue, and its ability to enhance rinsing was evaluated quantitatively. As a result, this method can classify the invasion stage of the tumor based on its natural 3D boundary and better elucidate tumor metastasis on a large scale based on a mouse model. iHE was also compatible with blood vessel staining, implying that we might obtain H&E staining and blood vessel information for a single tissue simultaneously. Thus, our results provide a novel tool to study tumor metastasis and infiltration in a mouse model. Furthermore, ultrasound shows promise in facilitating other types of biological tissue processing, such as optical clearing, immunolabeling and chemical staining of intact tissues.
3.1 Wndchrm-based analysis of morphology in noncancerous and gastric cancer tissues
To quantitatively assess biological morphology of cell and tissue conditions, we performed a machine learning analysis using the wndchrm algorithm, and specific image measurements (Figure 1A). Image data-sets were constructed in accordance with pathological diagnosis, using tissue microarrays derived from human stomach adenocarcinoma patients. Fifty-four H&E-stained tissue images of 1360 × 1024 pixels were collected for each class: Noncancer, Grade 1 (well differentiated), Grade 2 (moderately differentiated), and Grade 3 (poorly differentiated) (Figure 1B, Figure S1A and Table S1). Briefly, wndchrm extracted image features from all images of each defined class, and trained a classifier to discriminate between the classes using training data-sets. The classification performance was then validated with test images that were randomly selected, where these steps were automatically performed. We carried out 20 cross-validation analyses among the noncancer and grades 1-3 of gastric cancer (Figure 1A, left). As an initial step of the analysis, we examined the optimal number of images necessary for efficient classification. The results showed that the value of classification accuracy (CA) improved with increasing numbers of training images (Figure S1B), while that of standard errors became smaller as often seen in machine-learning analyses. 31 The best classification was found with 54 training images at CA 0.78 (the maximum CA is possibly 1.0), and this CA value was markedly higher than random classification at CA 0.25. Furthermore, the relative similarities among the classes were visualized with dendrograms (Figure 1C). In addition, using 20 images in each classes, we confirmed the classification similarity between noncancer, chronic gastritis and grades 1-3 (Figure S1C). Performance of the classification test was sufficient, as its specificity and sensitivity to discriminate cancer grades from noncancerous tissues was 100% and 92%, respectively (Table S2, upper). For additional assessment of morphological features, we divided each class into two subclasses, and measured the degree of dissimilarities of grades 1-3 from noncancerous tissues, as indicated by morphological distance (MD) (Figure 1D). The MD from Non-cancer_1 showed similarity to Noncancer l_2 and dissimilarity to cancer tissues of three grades. Furthermore, when the images were digitally tiled (see Methods), the numbers of training images increased, but the overview of the tissues was lost. However, CA values were largely unchanged at 0.79-0.69 using these tiled images (Figure S1D), suggesting that local morphology as well as histological overview are indicators to discriminate between noncancer and cancer tissues.
We then performed detailed binary comparisons between noncancer samples and each grade of gastric cancer to evaluate the effectiveness of wndchrm using each of the 54 images that showed sufficient CA values (Figure S1E). ROC curve analyses verified the accuracy of classifications, because AUCs were 0.99, 0.98, and 0.99 for Noncancer versus Grades 1, 2, and 3, respectively (the maximum AUC is 1.0, in contrast to random assignment of 0.5) (Figure 1E). Representative lists of informative image features in each classification test were indicated according to relative Fisher discrimination scores (Figure 1F). Many sets of image features were commonly used to discriminate Grades 1, 2, and 3 from Non-cancer (r > 0.7), although some distinct features were also involved (data not shown). Our results showed that wndchrm analyses highly recapitulated the human-based pathological examinations of H&E images of cancer tissues.
3.2 Wndchrm-based analysis reveals informative features of H&E-stained images
To understand which morphological features contribute to classification of noncancer and cancer grades, we digitally deconvolved the H&E RGB (Red, Green, Blue) images into hematoxylin and eosin channels in gray scales (Figure 2A). 30 Cellular nuclei and cytoplasmic components are generally stained with hematoxylin and eosin, respectively. 4, 10 Using the deconvolved images for noncancer and grades 1-3, as shown in Figure 2B and Figure S2, we measured CA among noncancer and grades 1-3 (Figure 2C). Cross-validation tests of hematoxylin and eosin images indicated equivalent CA values (0.72 and 0.69, respectively). Sensitivity and specificity were equally high at 82%-98% (Table S2, second from upper), suggesting that hematoxylin and eosin images contain morphological features distinguishing between cancer and noncancerous tissues.
A typical list of informative image features in the classification test was created according to relative Fisher discrimination scores, and showed overall similarities (Figure 2D). Pearson correlation coefficient value was weak between hematoxylin and eosin images (r = 0.55), suggesting the presence of unique morphological characteristics in either image. Consistently, MDs from Non-cancer_1 in both hematoxylin and eosin images showed dissimilarity between noncancerous and cancer tissues to a similar extent (Figure 2E,F).Thus, wndchrm analyses implied the presence of informative features in hematoxylin and eosin-stained images of cancer tissues.
3.3 Characterization of nuclear morphology in gastric cancer tissues
Our classification analysis of hematoxylin-stained images indicated that nuclear morphologies are distinct in noncancer and gastric cancers (grades 1-3), as shown by CA values (Figure 2C). To assess nuclear morphology, we measured two characteristics of the nucleus area and total intensity (Figure 1A, right). Using an image analysis software (Cellomics CellInsight), each measurement region was detected with a fixed size of 1024 × 1024 pixels from original tissue images (Figure 3A). By counting >12 000 nuclei, we found that nuclear area was significantly larger in cancer tissues, compared to noncancerous tissues (Figure 3B), and that signal intensity was also higher in cancer cells (Figure 3C). Because nuclei were densely distributed and sometimes overlapping in cancer tissues, probably due to high growth activities, we then attempted to measure this feature, using the nuclear area that was continuously stained with hematoxylin. We set the software to recognize the hematoxylin-positive area which was larger than the defined threshold (13 200 pixels), as shown in Figure 3D. The area with clustered nuclei was present prominently in Grades 1 and 2 of gastric cancers, but scarcely in Grade 3 (Figure 3E,F). Summary statistics for the area and total intensity are shown in Table S3, indicating that nuclear morphology is an advantageous parameter for cancer classification.
3.4 Expression levels of nuclear ATF7IP/MCAF1 are correlated with H&E images
It has been reported that various nuclear factors 37, 38 and membrane/soluble factors 39, 40 are involved in morphology of cells and tissues. We next investigated biological links between molecular expression and morphological features in gastric cancer tissues, using molecular marker-based analysis or fact-driven analysis.
To examine how H&E images can be classified based on molecular expression, we chose two cancer-associated proteins: nuclear ATF7IP/MCAF1 and membranous PD-L1 (Figures 4 and 5). ATF7IP/MCAF1 is an epigenetic factor involved in heterochromatin formation and gene regulation, which is frequently overexpressed in various kinds of tumors including gastric cancers. ATF7IP/MCAF1 functions for either DNA methylation-based gene repression or the transcription factor Sp1-mediated gene activation. 36 On the other hand, PD-L1 is generally produced by cancer cells to escape immune surveillance, and is a molecular target for cancer immune therapy. 41-43 Previous report showed that the PD-L1gene promoter is regulated by DNA methylation or Sp1 binding in cancer cells. 44, 45 There is the possibility that ATF7IP/MCAF1 may control PD-L1 expression via Sp1,as indicated by published ChIP-seq data of colon cancer (Figure S3A).
We performed both H&E staining and IHC using serial sections of tissue (Table S4). After the section slices were made from a paraffin block and stained, we carefully aligned the H&E with IHC images manually (Figure S3B-D). We selected 32 sites from H&E images (each 1360 × 1024 pixels) and the corresponding IHC images for ATF7IP/MCAF1 expression. Gastric cancer tissue and adjacent noncancerous regions showed high and low expression of ATF7IP/MCAF1, respectively (Figure 4A, Figure S3E,F). The levels of IHC signals were confirmed by quantification of their signals (Figure 4B). Based on the expression levels of ATF7IP/MCAF1, we then classified H&E images using wndchrm to low and high expression of this protein (CA 0.95-1.00, which shows high accuracy, regardless of image numbers) (Figure S3G), suggesting that gastric cancer tissues as tested can be clearly divided to these two classes. In addition, sensitivity and specificity of ATF7IP/MCAF1 signals were 100% and 98%, respectively (Table S2,second from lower). To evaluate the CA between low and high classes of ATF7IP/MCAF1, we arranged subclasses in H&E images (Low 1, Low 2, High 1 and High 2). Low 2 had similarity with Low 1, but significant difference with High 1 (Figure 4C). In addition, alignment of relative Fisher scores indicated weak correlation between the two comparisons (Figure 4D, r = 0.44), suggesting the presence of feature differences. Moreover, each MD from Low 1 in the feature space and the dendrogram showed morphological dissimilarity between low and high classes of ATF7IP/MCAF1 (Figure 4E,F).These suggested that expression levels of this protein are correlated with tissue morphology.
3.5 Expression levels of cytoplasmic PD-L1 are correlated with H&E images
We further investigated whether expression of the membranous protein PD-L1 in cancer is linked to tissue morphology. We again performed H&E staining and IHC with anti-PD-L1 antibodies, using serial sections of tissue microarrays in gastric cancer samples (Figure 5A and Table S5). We quantified IHC signal levels of PD-L1 staining, grouped into low and high expression of this protein, and further created data-sets of the corresponding H&E image (1360 × 1024 pixels) (Figure 5B, Figure S4A,B). Furthermore, we evaluated the tumor proportion score (TPS) by counting positively stained cells in 100 cells per image and found that PD-L1 High showed significantly higher TPS, while PD-L1 Low had very low TPS(Figure S4C).The H&E images were classified as Low and High PD-L1, at CA 0.86 using 60 images (Figure S4D). Sensitivity and specificity of PD-L1 signals were 88% and 84%, respectively (Table S2, lower). We confirmed the morphological dissimilarity between PD-L1 Low and High subclasses as shown by the CA (Figure 5C).The relative Fisher discriminant scores of image features suggested the presence of features responsible for the dissimilarity (Figure 5D,r = 0.59). Each MD from Low 1 in the feature space and the dendrogram showed morphological dissimilarities between Low and High classes of PD-L1 (Figure 5E,F).
Collectively, these results indicated that the expression of ATF7IP/MCAF1 and PD-L1 is correlated with tissue characteristics, suggesting that the spatial appearance of the cancer-associated proteins reflects morphological information of the pathological tissues.
Classification of Hematoxylin and Eosin-Stained Breast Cancer Histology Microscopy Images Using Transfer Learning with EfficientNets
Breast cancer is a fatal disease and is a leading cause of death in women worldwide. The process of diagnosis based on biopsy tissue is nontrivial, time-consuming, and prone to human error, and there may be conflict about the final diagnosis due to interobserver variability. Computer-aided diagnosis systems have been designed and implemented to combat these issues. These systems contribute significantly to increasing the efficiency and accuracy and reducing the cost of diagnosis. Moreover, these systems must perform better so that their determined diagnosis can be more reliable. This research investigates the application of the EfficientNet architecture for the classification of hematoxylin and eosin-stained breast cancer histology images provided by the ICIAR2018 dataset. Specifically, seven EfficientNets were fine-tuned and evaluated on their ability to classify images into four classes: normal, benign, in situ carcinoma, and invasive carcinoma. Moreover, two standard stain normalization techniques, Reinhard and Macenko, were observed to measure the impact of stain normalization on performance. The outcome of this approach reveals that the EfficientNet-B2 model yielded an accuracy and sensitivity of 98.33% using Reinhard stain normalization method on the training images and an accuracy and sensitivity of 96.67% using the Macenko stain normalization method. These satisfactory results indicate that transferring generic features from natural images to medical images through fine-tuning on EfficientNets can achieve satisfactory results.
1. Introduction and Background
One of the leading causes of death in women throughout the world is breast cancer . It is defined as a group of diseases in which cells within the tissue of the breast alter and divide in an uncontrolled manner, generally resulting in lumps or growths. This type of cancer often begins in the milk glands or ducts connecting these glands to the nipple. In the beginning stages of the illness, the small tumour that appears is much easier to treat effectively, averting the disease’s progression and decreasing the morbidity rates this is why screening is crucial for early detection .
The process of breast cancer diagnosis begins with palpation, periodic mammography, and ultrasonic imaging inspection. The results of these procedures indicate whether further testing is required. If cancer is suspected in a patient, a biopsy is performed and tissue for microscopic analysis is procured so that a pathologist may conduct a histological examination of the extracted tissue to confirm the diagnosis [2, 3]. Once the biopsy is complete, the tissue is analyzed in a laboratory. The tissue preparation process must begin with formalin fixation and, after that, embedding in paraffin sections. The paraffin blocks are then sliced and fixed on glass slides. Unfortunately, interesting structures such as the cytoplasm and nuclei in the tissue are not yet apparent at this point. The lack of clarity in the tissue necessitates staining of the tissue so that the structures can become more visible. Typically, a standard and well-known staining protocol, using hematoxylin and eosin, is applied. When added to the tissue, the hematoxylin can bind itself to deoxyribonucleic acid, which results in the nuclei in the tissue being dyed a blue/purple color. On the other hand, the eosin can bind itself to proteins, and, as a result, other relevant structures such as the stroma and cytoplasm are dyed a pink color. Traditionally, after staining, the glass slide is coverslipped and forwarded to a pathologist for examination . Routinely, the expert gathers information on the texture, size, shape, organization, interactions, and spatial arrangements of the nuclei. Additionally, the variability within, density of, and overall structure of the tissue is analyzed. In particular, the information concerning the nuclei features is relevant for distinguishing between noncarcinoma and carcinoma cells. In contrast, the information concerning the tissue structure is relevant for distinguishing between in situ and invasive carcinoma cells .
The noncarcinoma class consists of normal tissue and benign lesions these tissues are nonmalignant and do not require immediate medical attention. In situ and invasive carcinoma, on the other hand, are malignant and become continuously more lethal without treatment. Specifically, in situ carcinoma refers to the presence of atypical cells that are confined to the layer of tissue in the breast from which it stemmed. Invasive carcinoma refers to the presence of atypical cells that invades the surrounding normal tissue, beyond the glands or ducts from where the cells originated . Invasive carcinoma is complicated to treat, as it poses a risk to the entire body . This threat means that the odds of surviving this level of cancer decreases as the progression stages increase. Moreover, without proper and adequate treatment, a patient’s in situ carcinoma tissue can develop into invasive carcinoma tissue. Therefore, it is of paramount importance that biopsy tissue is examined correctly and efficiently so that a diagnosis can be confirmed and, subsequently, treatment can begin. Examples of histology images belonging to each of these classes are shown in Figure 1.
The task of performing a practical examination on the tissue is not simple and straightforward. On the contrary, it is rather time-consuming and, above all, prone to human error. The average diagnostic accuracy between professionals is around 75% . These issues can result in severe and fatal consequences for patients who are incorrectly diagnosed .
The advancement of image acquisition devices that create whole slide images (WSI) from scanning conventional glass slides has promoted digital pathology . The field of digital pathology focuses on bringing improvement in accuracy and efficiency to the pathology practice  by associating histopathological analysis with the study of WSI .
An excellent solution to address the limitations of human diagnosis is computer-aided diagnosis (CAD) systems, which are developed to automatically analyze the WSI and provide a potential diagnosis based on the image. These systems currently contribute to improving efficiency and reducing both the cost of diagnosis and interobserver variability [5, 10]. Even though current CAD systems that operate at high sensitivity provide relatively good performance, they will remain a second-opinion clinical procedure until the performance is significantly improved .
Recently, deep learning approaches to the development of CAD systems have produced promising results. Previous attempts to classify breast cancer histology images using a combination of handcrafted feature extraction methods and traditional machine learning algorithms required additional knowledge and were time-consuming to develop. Conversely, deep learning methods automate this process. These systems allow pathologists to focus on difficult diagnosis cases .
Hence, to ensure early diagnosis in breast cancer candidates, increase treatment success, and lower mortality rates, early detection is imperative. Although the advent of and advancements in computer-aided systems have benefited the medical field, there is plenty of room for improvement.
1.1. Research Problem
In general, the shortage of available medical experts , the time-consuming quest to reach a final decision on a diagnosis, and the issue of interobserver variability justify the need for a system that can automatically and accurately classify breast cancer histopathology images. Previous approaches to this problem have been relatively successful considering the available data and return adequate classification accuracies but tend to be computationally expensive. Thus, this work will explore the use of seven lightweight architectures within the EfficientNet family . Since the EfficientNet models were designed to optimize available resources, while maintaining high accuracies, a CAD system that performs at the level of the current state-of-the-art deep learning approaches, while consuming less space and training time, is desirable. Transfer learning techniques have become a popular addition to deep learning solutions for classification tasks. In particular, many state-of-the-art approaches utilize fine-tuning to enhance performance . Therefore, this research explores the application of seven pretrained EfficientNets for the classification of breast cancer histology images. Furthermore, the addition of stain normalization to the preprocessing step will be evaluated. Hence, the primary question that this research will answer is, “Can fine-tuned EfficientNets achieve similar results to current state-of-the-art approaches for the application of classifying breast cancer histology images?”
1.2. Research Contributions
In this research, the application of seven versions of EfficientNets with transfer learning for breast cancer histology image classification is investigated. The proposed architecture was able to effectively extract and learn the global features in an image, such as the tissue and nuclei organization. Of the seven models tested, the EfficientNet-B2 architecture produced superior results with an accuracy of 98.33% and sensitivity of 98.44%.
The key takeaway from this investigation is that the simple and straightforward approach to using EfficientNets for the classification of breast cancer histology images reduces training time while maintaining similar accuracies to previously proposed computationally expensive approaches.
1.3. Paper Structure
The remainder of the paper is structured as follows: Section 2, the literature review, provides details on previous successful approaches. Section 3, the methods and techniques, provides insight into the framework followed in this study. Section 4, the results, provides details of the results that were obtained during the research. Finally, section 5 elaborates on the insights of this work and concludes the paper.
2. Literature Review
Currently, computer-aided diagnosis (CAD) systems occupy the position of aiding physicians during the process of diagnosis, by easing their workload and reducing the disagreement that stems from the subjective interpretation of pathologists. However, the performance of these systems must be enhanced before they can be considered more dependable than a second-opinion system .
2.1. Traditional Approaches
In the traditional approach, expert domain knowledge is required so that the correct features may be handcrafted this is a time-consuming endeavor. Nevertheless, the approach yields acceptable results on the datasets used. For instance, Kowal  used multiple clustering algorithms to achieve nuclei segmentation on microscopic images. Segmentation made it possible to extract microscopic, textural, and topological features so that classifiers could be trained and images could be classified as either benign or malignant. The accuracy of patient-wise classification was in the range of 96–100%. It is worth noting that this method performs poorly when an image contains overlapping nuclei or a small number of nuclei. In this case, either the approach fails to identify the nuclei or the clustering algorithms could return unreliable results. Therefore, in order to attain an acceptable detection accuracy, a large number of sample images are required. Hence, it is evident that accurate nuclei segmentation is not a straightforward task this can also be attributed to the variability in tissue appearance or the presence of clustered or tightly clumped nuclei .
An alternative approach is utilizing information on tissue organization as in the work by Belsare et al. , which presents a framework to classify images into malignant and nonmalignant. Firstly, segmentation was done using spatio-color-texture graphs. After that, statistical feature analysis was employed, and classification was achieved with a linear discriminant classifier. The choice of this classifier considerably impacted the outcome of this approach, as the result outperformed the use of k-nearest-neighbor and state vector machine classifiers, especially for the detection of nonmalignant tissue. Accuracies of 100% and 80% were achieved for nonmalignant and malignant images, respectively.
2.2. Deep Learning Approaches
The increase in the availability of computing power has led to the emergence of advanced architectures called convolutional neural networks (CNNs). Contrary to the conventional approach, no expert domain knowledge is required to define algorithms for segmentation, feature extraction, and classification, but instead expert knowledge is needed to annotate the dataset for a CNN to achieve superior results. Instead, these networks can automatically determine and extract discriminative features in an image that contribute to the classification of the image. Generally, a CNN will use a training set of images to learn features that are unique to each class so that when a similar feature is detected in an unseen image, the network will be able to assign the image to a class with confidence.
2.2.1. Convolutional Neural Network Approaches
The success of convolutional neural networks (CNNs) with general computer vision tasks motivated researchers to employ these models for classifying histopathology images. For the classification of hematoxylin and eosin-stained breast cancer histology images, both Araújo et al.  and Vo et al.  used the Bioimaging 2015 dataset  and classified the images into four classes (normal, benign, in situ, and invasive) and two groups (carcinoma and noncarcinoma). The former work  proposes a CNN that can integrate information from multiple histological scales. The process begins with stain normalization via the method proposed by Macenko et al.  in a bid to correct color discrepancies. After that, 12
overlapping patches were extracted from each image. The chosen size of the patches ensures that no relevant information is lost during extraction and, therefore, every patch can be appropriately labeled. Then, data augmentation was used to increase the number of images in the dataset. Finally, a patch-wise trained CNN and a fusion of a CNN and support vector machine classifier (CNN + SVM) were used to determine the patch class probability. Image-wise classification was attained through a patch probability fusion method. The evaluation showed that using majority voting strategy as the fusion method produced the best results. Considering all four classes, patch-wise classification with the CNN achieved an accuracy rate of 66.7%, while the CNN + SVM achieved an accuracy rate of 65%. The image-wise classification achieved higher results at 77.8% accuracy for both classifiers. With only two classes, patch-wise accuracy for the CNN was 77.6%, and for the CNN + SVM, the approach yielded 76.9% accuracy. The image-wise classification for the 2-class task produced the best results at 80.6% for the CNN and 83.3% for the CNN + SVM. The reason for the lower patch-wise classification is that images may contain sections of normal-looking tissue. Since during patch generation the extracted patches inherit the image’s label, this may confuse the CNN. The increase in image-wise classification accuracy is due to the fusion method that is applied. The authors also recorded the sensitivity rates for each of the classes. It is worth noting that overall, for image-wise classification, the approach was more sensitive to the carcinoma class than the noncarcinoma class. This outcome, although not ideal, is preferable since the architecture that was proposed focuses on correctly classifying the carcinoma (malignant) instances .
The approach taken by Vo and Nguyen  proposed a combination of an ensemble of deep CNNs and gradient boosting tree classifiers (GBTCs). Stain normalization via Macenko et al.  and data augmentation were the initial steps of the process. Unlike the standard data augmentation method of rotating and flipping images, the proposed method  incorporates reflection, translation, and random cropping of the images. The normalized and augmented data was then used to train the proposed architecture. Specifically, three deep CNNs (Inception-ResNet-v2) were trained using three different input sizes:
. Then, visual features were extracted and fed into GBTCs, which increased classification performance. The majority voting strategy was used to merge the outputs of the GBTCs, resulting in a much more robust solution. Recognition rates of 96.4% for the 4-class classification and 99.5% for the 2-class classification were reported. This result surpasses state-of-the-art achievements. An interesting note is that the authors added global average pooling layers in place of dense (fully connected) layers, and this did not negatively impact the accuracy of the ensemble. Similar to Araújo et al. , the authors of this work recorded the sensitivities of their approach. The results indicate that, for the 4-class task, the proposed method struggles with the classification of the in situ instances, while the other three classes have incredibly high sensitivities. For the 2-class task, the approach yields a 100% sensitivity on carcinoma instances and 98.9% on noncarcinoma instances. These results indicate that the approach was able to successfully learn both local and global features for the multiclass and binary classification. However, the downfall of this approach is the computational expense.
2.2.2. Convolutional Neural Network with Transfer Learning Approaches
For the TK-AlexNet proposed by Nawaz et al.  to classify breast cancer histology images, the classification layers of the AlexNet architecture were replaced with a single convolutional layer, and a max-pooling layer was added before the three fully connected layers with 256, 100, and 4 neurons, respectively. The input size of the proposed network  was increased to 512 512. The transfer learning technique used in this application was to fine-tune the last three layers on the ICIAR2018 dataset after having the entire network trained on the ImageNet dataset. The images were stain-normalized with the method proposed in Macenko et al. . An interesting fact is that the authors compared the performance of the model with both non-stain-normalized and stain-normalized images and concluded that using the latter resulted in a gain in performance. After that, data augmentation techniques such as mirroring and rotation were applied, and overlapping patches of size were extracted from each image. Hence, there was a total of 38400 images generated. Evaluation of the model was done using a train-test split of 80%–20%.
The image-wise accuracy reported in  was 81.25%, and the patch-wise accuracy was 75.73%. A noteworthy observation is that the normal and benign classes were classified with 85% sensitivity however, the in situ and invasive carcinoma classes were classified with 75% sensitivity. For a model to be practical as a second-opinion system, it should ideally have a higher sensitivity to the carcinoma class given the dangers of misdiagnosis.
For this classification task, the Inception-ResNet-v2 was used by Ferreria et al. . The classification layers of the base model were replaced by a global average pooling layer, a dense (or fully connected) layer with 256 neurons, a dropout layer with a dropout rate of 0.5, and a final dense layer of 4 neurons. Moreover, the input size of the network was changed to . Reshaping the images does not significantly impact the form of the cellular structures however, it does reduce computational cost . The authors did not incorporate stain normalization into their experiments. Data augmentation techniques such as image flips (horizontal and vertical), a 10% zoom range, and shifts (horizontal and vertical) were used to increase the dataset. These particular techniques were chosen with care because if the augmentation causes too much distortion, the anatomical structures in the image could be destroyed , and this may result in the network having difficulty extracting discriminative features during training.
Two forms of transfer learning were used in this experiment. At first, only the dense (fully connected) layers of the model were trained. This technique is referred to as feature extraction since the network is using pretrained features (from ImageNet) to classify the breast cancer histology images. The result of this step is that only the weights of the dense layers were adjusted. This aids in overfitting . Afterwards, a certain number of layers were unfrozen so that the network could be fine-tuned. Early stopping with a patience of 20 epochs, and a checkpoint callback monitoring minimum validation loss were the additional techniques implemented to avoid overfitting. The dataset was randomly split into 70% training, 20% validation, and 10% testing. The test set achieved an accuracy of 90%.
In a study by Kassani et al. , five different architectures (Inception-v3, Inception-ResNet-v2, Xception, VGG16, and VGG19) were investigated for the classification of the ICIAR2018 dataset. Two stain normalization methods were observed in this study: Macenko et al.  and Reinhard et al. . Data augmentation included vertical flips, contrast adjustment, rotation, and brightness correction. The data was split into 75% and 25% for training and testing, respectively. The images were resized to pixels with the help of bicubic interpolation. For each of the models, features were extracted from specific blocks, particularly the layer after a max-pooling layer. The extracted features were put through a global average pooling layer and then concatenated to form a feature vector which was fed into an MLP (multilayer perceptron) set with 256 neurons for final classification. Of these models, the modified Xception network trained with Reinhard stain-normalized images performed the best, with a reported accuracy score of 94%. Overall, the Xception architecture performed the best for both of the stain normalization methods, and the Reinhard  technique produced higher accuracies than Macenko . The other architectures ranked in the following order: Inception-v3, Inception-ResNet-v2, VGG16, and VGG19. Interestingly, the approximate parameters for these architectures are 23 million, 54 million, 138 million, and 143 million, respectively. One could hypothesize that an increase in parameter count translates to a decrease in accuracy of this dataset. This indicates that the bigger architectures may have more difficulty extracting critical features from training images, even if measures are taken to enlarge the dataset being used. The results of this study also emphasize the benefit of incorporating stain normalization into preprocessing and how choosing the correct method improves accuracy significantly. Table 1 shows a comparison summary of related deep learning techniques in the literature.
Histopathology-driven artificial intelligence predicts TMB-H colorectal cancer
IMAGE: A representative microscopic image of the hematoxylin and eosin (H&E) stained tumor mutational burden-high colorectal cancer tumor. Digital information from such this neoplastic and also non-neoplastic images is transformed and. view more
Credit: Niigata University
Niigata, Japan - Biomarkers are important determinants of appropriate and effective therapeutic approaches for various diseases including cancer. There is ample evidence pointing toward the significance of immune check point inhibitors (ICI) against cancer, and they showed promising clinical benefits to a specific group of patients with colorectal cancer (CRC). Several reports demonstrated the efficacy of biomarkers such as programmed death-1 protein ligand (PD-L1), density of tumor-infiltrating lymphocytes (TILs), and tumor mutational burden (TMB), to determine the patient responsiveness for the efficient use of ICIs as therapeutics against cancer.
A high level of TMB (TMB-H), which reflects elevated total number of non-synonymous somatic mutations per coding area of a tumor genome and normally derived from gene panel testing, is recognized as a promising biomarker for the ICI therapies of various solid cancers. However, in clinical practice, it is not feasible to perform gene panel testing for all cancer patients.
In addition, the studies of Dr. Shimada group also provided means to predict TMB-H CRC only by using the TIL information from the H&E slides from the patients' tumor tissues. However, considering that the patients in the studied cohort were not treated with any ICIs, no conclusions could be drawn regarding their ICI responsiveness following the TMB-H diagnosis and it was suggested that future clinical trials need to be conducted to address whether TIL alone can be useful as a predictive biomarker for the efficacy of ICIs. Dr. Shimada says about the present study: "We have developed artificial intelligence to predict genetic alterations in colorectal cancer by deep learning using hematoxylin and eosin slides. This artificial intelligence is important in solving the cost problems associated with genetic analysis and facilitating personalized medicine in colorectal cancer."
Overall, the studies by Dr. Shimada and associates provide a cost and time effective and reliable method to inform the clinicians if the CRC patient they are managing can benefit from Immune Checkpoint Inhibitor (including inhibitors of the PD-1 protein and its ligand, PD-L1) therapy, without implicating the use of gene panel.
Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.
HER2 Molecular Marker Scoring Using Transfer Learning and Decision Level Fusion
In prognostic evaluation of breast cancer, immunohistochemical (IHC) marker human epidermal growth factor receptor 2 (HER2) is used for prognostic evaluation. Accurate assessment of HER2-stained tissue sample is essential in therapeutic decision making for the patients. In regular clinical settings, expert pathologists assess the HER2-stained tissue slide under microscope for manual scoring based on prior experience. Manual scoring is time consuming, tedious, and often prone to inter-observer variation among group of pathologists. With the recent advancement in the area of computer vision and deep learning, medical image analysis has got significant attention. A number of deep learning architectures have been proposed for classification of different image groups. These networks are also used for transfer learning to classify other image classes. In the presented study, a number of transfer learning architectures are used for HER2 scoring. Five pre-trained architectures viz. VGG16, VGG19, ResNet50, MobileNetV2, and NASNetMobile with decimating the fully connected layers to get 3-class classification have been used for the comparative assessment of the networks as well as further scoring of stained tissue sample image based on statistical voting using mode operator. HER2 Challenge dataset from Warwick University is used in this study. A total of 2130 image patches were extracted to generate the training dataset from 300 training images corresponding to 30 training cases. The output model is then tested on 800 new test image patches from 100 test images acquired from 10 test cases (different from training cases) to report the outcome results. The transfer learning models have shown significant accuracy with VGG19 showing the best accuracy for the test images. The accuracy is found to be 93%, which increases to 98% on the image-based scoring using statistical voting mechanism. The output shows a capable quantification pipeline in automated HER2 score generation.
One limitation of percutaneous image-guided ablation when compared with surgery is the lack of pathologic evidence that the target lesion was completely ablated with sufficient tumor-negative margins. Most studies indicate that positive surgical margins have been associated with a higher risk of local recurrence and shorter overall survival (28,29), a conclusion that has been verified in recent studies performed in the context of modern preoperative chemotherapy (30,31).
Similarly, after RF ablation, the minimal ablation margin is a factor associated with local tumor control (23,32–35). Treatment effectiveness is routinely assessed with postprocedural imaging (US, CT, or MR imaging) to evaluate if the intended ablation margin has been reached, and if it has not, to direct further treatment (23,36,37). Several methods have been proposed for precise calculation of the ablation margin, including fusion imaging (32) and use of anatomic intrahepatic landmarks (23). All of these methods suffer from differences when comparing images from scans performed at different times. Despite progress in fusion and registration software, limitations of these techniques still present a challenge when attempting accurate calculation of the margin of the ablation zone around a previously treated tumor (23,32). Despite the use of these imaging modalities, incomplete tumor treatment and LTP remain common after RF ablation (7–10).
The sole use of imaging in the assessment of ablation margins presumes that the entire depicted ablation zone contains necrotic or nonviable cells however, few studies have validated image findings with tissue characteristics after ablation. In the early radiologic-pathologic correlation study performed by Goldberg et al (38), imaging failed to depict peripheral residual untreated tumor foci, even in four of five cases in which initial tumor and ablation zone sizes were identical. Furthermore, (a) prior studies performed to evaluate tissue collected from the electrodes used for ablation and (b) our study, in which we preformed direct assessment of the ablation zone with biopsy, show that viable tumor cells may be present within the ablation zone, even when postprocedural imaging displays sufficient ablation margins and findings compatible with radiographic technical success (12,13,16). It appears that these incompletely treated viable tumor cells within the ablation zone “escape” the spatial resolution of available anatomic and functional imaging modalities. In addition, the assessment of tissue examinations in the ablation zone introduces an objective tool with which to assess ablation effectiveness. This assessment is less vulnerable to operator variability and technical limitations, such as those described in the assessment and calculation of the ablation margin (23), and it may be easier to reproduce. Nevertheless, biopsies and sampling errors can also occur, and residual tumor may not be detected in some cases. This can explain the few recurrences seen in our study even after negative biopsy results and sufficient margins were obtained.
Histopathologic examination of tumors excised immediately after RF ablation has shown regions of altered cellular morphology within the treated volume that do not correspond to coagulative necrosis or classic tumor cell appearance (38). These areas may represent either (a) cells in the early stages of irreversible apoptosis or coagulative necrosis or (b) viable cells maintaining their proliferative potential, allowing them to replicate once the cellular insult is removed (21,22). Thus, it is prudent to further interrogate such cells with available immunohistochemical staining for cytosolic and mitochondrial enzyme activity, especially in the immediate postprocedural setting. Days after RF ablation, the evolution of intracellular cascades leads to more pronounced cellular changes in the direction of irreversible apoptosis or necrosis that can be evaluated with standard hematoxylin-eosin staining (38,39). However, immunohistochemical staining has been considered superior to hematoxylin-eosin staining in the diagnosis of irreversible cellular damage up to 24 weeks after RF ablation, as noted in an earlier study by Morimoto et al (39).
An interesting and unique finding of our study is that immunohistochemistry did not alter the initial classification based on the interpretation of the morphologic (hematoxylin-eosin) stain. All 16 ablation zones containing tumor cells were also positive for Ki-67, OXP, or both. This observation differs from observations in previous studies in which ablated tissue was assessed and in which immunohistochemistry was considered necessary to determine viability and changed specimen classification in two (13%) of 15 cases (12,13). It is possible that this finding represents the acquired experience of study pathologists at our institution in the evaluation of ablated tissue, and it potentially justifies the future investigation of the utility of immediate postablation frozen sections and morphologic assessment to document complete tumor cell necrosis.
Viability of tissue adherent to electrodes was analyzed in a study by Snoeren et al (16), who used the autofluorescence method and glucose-6-phosphate diaphorase staining. They concluded that viable tissue was an independent risk factor for LTP . Despite the different methods used to document viability, the incidence of viable tissue in the Snoeren et al study (29.2%) was slightly greater than that in our study (24%) and in the prior studies by Sofocleous et al (19.1%) (12,13). This observation, in combination with the fact that viability in all studies was an independent predictor of LTP , may indicate that both methods of tissue evaluation possess similar efficacy.
Our study had several limitations. The most important factor limiting the weight of our conclusions was the relatively small number of enrolled patients (n = 47). In addition, pathologic evaluation with biopsies from the center and the margin of the ablation zone does not yield information about necrosis within the entire treated lesion volume, as opposed to the evaluation of resected tumors and their surgical margins. Moreover, specimens were classified as Ki-67 positive even if the evaluated tissue was only slightly positive at immunostaining. Estimation of the labeling percentage of the target CLM was not performed, thereby precluding investigation of any potential correlation between the level and grade of proliferation (Ki-67 positivity) and time to LTP . Another limitation of our study was that the minimal margin analyzed as a predictor for time to LTP was evaluated by using 4–8-week postablation CT images, while the presumed minimal margin targeted with biopsy was estimated by using CT images obtained immediately after ablation. That was a rough estimate suited to the time limitations of the procedure however, for this study, accurate estimation and recording of the minimal margin (using CT landmarks) was performed by using CT images obtained 4–8 weeks after RF ablation, as previously described in detail (23). Software capabilities with fusion of the preablation tumor with the ablation zone are currently under development and are evolving. This will enable accurate calculation of the margin during or immediately after ablation and at subsequent time points. As indicated in our work, a positive viable tumor biopsy had a strong and significant prognostic value, as it was an early biomarker of local tumor recurrence and ablation failure. The addition of biopsy of the ablation zone introduces a more objective and reproducible significant predictive assessment of the ablation zone rather than relying on the imaging findings and margin assessment alone.
Advances in Knowledge
■ Ablated tumors with posttreatment biopsies containing tumor cells positive for Ki-67 or OxPhos antibodies are 3.4 times more likely to recur (P = .008).