Addressing generalizability and clinical utility of AI-enabled virtual-IHC for melanocytic cells

Summary: Abstract Body: The use of histologically stained slides has long had an important role for diagnoses, with immunohistochemical (IHC) stains used to confirm important cell types and regions of interest needed for diagnosis. However, the material and monetary cost to regularly produce these stains, when necessary, can be costly or difficult for cases with limited sample availability. We developed a virtual IHC method using a deep learning model to perform segmentation of melanocytes from paired adjacent Melan-A IHC and H&E whole slide images (WSI) for the diagnosis of melanoma, as proof of principle. We developed a pipeline in which image processing was used to generate melanocytic masks from the paired slides without human annotation and then trained and evaluated on 23 WSI from a range of diagnostic classes (benign\dysplastic nevi, melanoma-in situ and malignant melanoma), achieving a mean intersection-over-union (mIoU) of 0.65. The robustness of the model was then tested against domain shifts from a different scanner (23 WSI), batch (49 WSI) and negative control vitiligo cases (23 WSI). We found that although these datasets showed initial decreased performance, we could recover both mIoU and visual performance with minimal fine-tuning of the model or the addition of out of domain data for the case of vitiligo. In a test of clinical utility, we also found no significant difference in diagnostic accuracies when a panel of 3 dermatopathologists were asked to make diagnoses with the actual IHC WSI compared with virtual IHC WSI. Together we find that the use of virtual IHC can closely match that of a real IHC and that with fine-tuning the model can be robust to different domain shifts. Mikio Tada¹, Rodrigo Torres¹, Ursula Lang^{2, 3}, Rony Francois^{2, 3}, Iwei Yeh^{2, 3}, Elizabeth Keiser⁴, Tatsiana Pukhalskaya³, Erin Amerson², Michael Keiser⁵, Maria L. Wei^{2, 1} 1. San Francisco VA Health Care System, San Francisco, CA, United States. 2. Dermatology, University of California San Francisco, San Francisco, CA, United States. 3. Pathology, University of California San Francisco, San Francisco, CA, United States. 4. Pathology, San Diego Veterans Health Care System, San Diego, CA, United States. 5. Institute for Neurodegenerative Diseases, University of California San Francisco, San Francisco, CA, United States. Bioinformatics, Computational Biology, and Imaging