Introduction
Immunohistochemistry (IHC) is a critical tool for tumor diagnosis and treatment. However, IHC is time consuming and costly, relies on the pathologists’ subjective interpretation. Deep learning has gained ground for predicting the biomarker expression on hematoxylin and eosin (H&E) staining images. In this study, we aim to create an automatic pipeline for the construction of deep learning models to generate virtual IHC output using H&E whole slide images and validate its clinical applicability by reader study.
Aims & Methods
We obtained 134 whole-slide images (WSIs) from 73 patients at Renmin Hospital of Wuhan University and extracted 415,463 patches from H&E-stained WSIs. Our study focused on five widely used biomarkers in pathology reports, namely anti-P40, CK-pan, Desmin, P53, and Ki-67, as recommended by the NCCN guidelines. We validated these models by using independent whole-slide images (WSIs). The number of WSIs used for each model were as follows: model-P40 (6 WSIs, 38389 Patches), model-PCK (9 WSIs, 72443 Patches), model-Ki67 (7 WSIs, 53162 Patches), model-Desmin (5 WSIs, 23974 Patches), and model-P53 (5 WSIs, 36822 Patches). We aimed to predict these biomarkers expression in H&E WSIs and quantify the percentage of Ki-67. There are three main steps: Firstly, we employed HEMnet neural network to align the corresponding IHC and H&E WSIs, transfering the molecular labels from IHC to H&E images. Secondly, pathologists reviewed the annotations made by HEMnet and corrected HE-IHC mismatched tiles using VIA labelling platform. Lastly, the deep learning models were trained to predict positive tiles. The performance of models was assessed by sensitivity, specificity, accuracy, and the area under the receiver operating characteristic curve (AUC). To assess the applicability of models, we randomly selected 30 cases including H&E and IHC WSIs for three pathologists to write the reports in the reader study by reading the interpretation of real IHC and AI-generated IHC.
![]()
![]()
Results
The sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) of the models for classifying positive pixels of each biomarker was: P40 (91.64%, 87.39%, 88.66%, AUC: 0.95), PAN-CK (92.40%, 89.35%, 90.81%, AUC: 0.96), Ki67 (89.63%, 86.04%, 86.77%, AUC: 0.95), Desmin (91.44%, 88.11%, 89.17%, AUC: 0.96), and P53 (75.44%,86.52%,83.04%, AUC: 0.90). The predicted Ki-67 percentage was comparable to the assessments made by expert pathologists, with a difference ranging from 7.00%±3.69%. In the reader study, compared with real-IHC pathology report, the accuracy of AI-IHC pathology report of each biomarker was: Desmin(100%,66/66), Pan-CK(98.3% , 59/60), P53(98.3% ,59/60), T-stage(84.8% .56/66), P40(96.7% ,29/30). and there is no significant difference between the quatification of Ki-67 score(t= 0.253 ,P= 0.800).
| The result of Reader Study |
|---|
| Items | Pathology report's accuracy of AI-IHC compared with Real-IHC | T test of Ki-67 |
| Desmin | Pan-CK | P53 | T-stage | P40 |
| Pathologist 1 | 100.0% (22/22) | 100.0% (20/20) | 70.0% (21/30) | 68.2% (15/22) | 100.0% (10/10) | t= 0.475 (P= 0.637) |
| Pathologist 2 | 100.0% (22/22) | 95.0% (19/20) | 66.7% (20/30) | 95.5% (21/22) | 100.0% (10/10) | t= - 0.775 (P= 0.443) |
| Pathologist 3 | 100.0% (22/22) | 100.0% (20/20) | 60.0% (18/30) | 90.9% (20/22) | 90.0% (9/10) | t= - 0.510 (P= 0.613) |
| Average | 100.0% (66/66) | 98.3% (59/60) | 98.3% (59/60) | 84.8% (56/66) | 96.7% (29/30) | t= 0.253 (P= 0.800) |
Conclusion
Our deep learning models have demonstrated promising performance in predicting gastrointestinal tumor biomarkers based on H&E slides. The utilization of these models can alleviate the burden of IHC staining and clinicopathological assessments. Furthermore, these models have the potential to aid in patient grading, facilitate informed treatment decisions, and provide valuable prognostic assessments.
Disclosure
There is no conflict of interest in this abstract.