Piloting a Deep Learning Model for Predicting Nuclear BAP1 Immunohistochemical Expression of Uveal Melanoma from Hematoxylin-and-Eosin Sections

Background: Uveal melanoma (UM) is the most common primary intraocular malignancy in adults. Monosomy 3 and mutation are strong prognostic factors predicting metastatic risk in UM. Nuclear BAP1 (nBAP1) expression is a close immunohistochemical surrogate for both genetic alterations. Not all laboratories perform routine BAP1 immunohistochemistry or genetic testing, and rely mainly on clinical information and anatomic/morphologic analyses for UM prognostication. The purpose of our study was to pilot deep learning (DL) techniques to predict nBAP1 expression on whole slide images (WSIs) of hematoxylin and eosin (H&E) stained UM sections.

Methods: One hundred forty H&E-stained UMs were scanned at 40 × magnification, using commercially available WSI image scanners. The training cohort comprised 66 BAP1 and 74 BAP1 UM, with known chromosome 3 status and clinical outcomes. Nonoverlapping areas of three different dimensions (512 × 512, 1024 × 1024, and 2048 × 2048 pixels) for comparison were extracted from tumor regions in each WSI, and were resized to 256 × 256 pixels. Deep convolutional neural networks (Resnet18 pre-trained on Imagenet) and auto-encoder-decoders (U-Net) were trained to predict nBAP1 expression of these patches. Trained models were tested on the patches cropped from a test cohort of WSIs of 16 BAP1 and 28 BAP1 UM cases.

Results: The trained model with best performance achieved area under the curve values of 0.90 for patches and 0.93 for slides on the test set.

Conclusions: Our results show the effectiveness of DL for predicting nBAP1 expression in UM on the basis of H&E sections only.

Translational Relevance: Our pilot demonstrates a high capacity of artificial intelligence-related techniques for automated prediction on the basis of histomorphology, and may be translatable into routine histology laboratories.