Application of cutaneous melanoma multiple instance learning model for conjunctival melanoma whole slide images

Summary: Abstract Body: While advancements in artificial intelligence have rapidly progressed, there are still many challenges to applying computer vision models to whole slide images (WSIs) in pathology due to their scattered artifacts and large size. Multiple instance learning (MIL) is a type of weakly supervised deep learning that addresses these challenges by dividing WSIs into smaller regions for analysis, while only utilizing slide-level labels. We established proof of concept that MIL models trained on cutaneous melanoma WSIs can be applied to conjunctival melanoma samples, which is particularly important given the smaller number of conjunctival samples. In prior research, we developed several weakly supervised deep learning MIL models (512x512 pixel tiles, extracted 1024-dimensional feature vectors) trained on cutaneous melanocytic lesions and achieved high performance in classifying cutaneous melanoma versus benign lesions. The highest performing model achieved an AUC of 0.96 ± 0.08 using 10-fold cross-validation, using only 82 samples for training, illustrating the applicability of MIL for small, high-resolution datasets. We established proof of concept in the effective application of the cutaneous melanoma models to conjunctival melanocytic lesions. The conjunctival dataset consisted of 53 WSIs across five diagnostic classes: nevus (27), low-grade conjunctival melanocytic intraepithelial lesions (CMIL) (7), high-grade CMIL (4), melanoma in situ (MIS) (6), and invasive melanoma (9). A cutaneous melanoma model performed well in distinguishing conjunctival melanoma versus nevi with an AUC of 0.90 with a sensitivity of 0.944 and a specificity of 0.644. These results demonstrate the generalizability of cutaneous melanoma weakly supervised deep learning models for conjunctival lesions, highlighting their potential to be adapted for diagnosing other lesion types. Mikio Tada¹, Alexandra So¹, Rodrigo Torres^{1, 2}, Iwei Yeh¹, Ursula Lang¹, Erin Amerson¹, Michael Keiser¹, Maria L. Wei^{1, 2} 1. University of California San Francisco, San Francisco, CA, United States. 2. San Francisco VA Health Care System, San Francisco, CA, United States. Pigmentation, Melanoma, and Melanoma Immune Surveillance