Utilizing self-supervised machine learning to improve risk stratification in cutaneous squamous cell carcinoma

Summary: Abstract Body: Staging systems for cutaneous squamous cell carcinoma (cSCC) often fail to reliably predict poor outcomes. Up to 40% of poor-outcome cases are classified within low-risk groups (BWH), while high-risk cSCC cases show a 46% upstaging rate1,2. We developed the rank-aware contextual reasoning (RACR) multiple instance learning (MIL) approach to predict metastasis risk and identify novel histological phenotypes3. Our approach combines self-supervised histological features with hierarchical graph networks to capture the tumor microenvironment and structural context. The model was trained on 307 patients and validated on 77 patients from Mayo Clinic, Arizona, and Complejo Asistencial Universitario de Salamanca, Spain. Metastatic risk was binary, defined as metastasis within 5 years of biopsy. The model achieved a mean AUC of 0.774 (±0.038) across 5-fold cross-validation, outperforming existing MIL methods by 3-5%. High-risk tissue regions identified by the model were clustered and validated by a board-certified pathologist, revealing histopathological clusters with key patterns and features: 1) Poor differentiation with sheet-like growth pattern and large tumor cells with prominent nucleoli 2) Tumor infiltration with desmoplastic stroma and fibrosis and tumor strands at tumor-stromal interface 3) Acellular benign structures, suggesting proximity to deeper structures and 4) Lymphocytic inflammation with variable differentiation. These findings demonstrate the potential of our model to enhance risk stratification in cSCC by identifying histopathologic features associated to metastasis. Anirudh Choudhary¹, Zachary Leibovit-Reiben², Alyssa L. Stockard², Angelina Hwang², Jacob Kechter², Krishnakant Saboo¹, Nneka Comfere³, Steven Nelson², Emma Johnson³, Leah Swanson², Olayemi Sokumbi⁴, Paul Arnold⁵, Javier Canueto⁶, David DiCaudo², Ravishankar Iyer¹, Aaron R. Mangold² 1. Electrical and Computer Engineering, University of Illinois Urbana-Champaign, Urbana, IL, United States. 2. Dermatology, Mayo Clinic Arizona, Scottsdale, AZ, United States. 3. Dermatology, Mayo Clinic Minnesota, Rochester, MN, United States. 4. Dermatology, Mayo Clinic in Florida, Jacksonville, FL, United States. 5. Carle Illinois College of Medicine, Urbana, IL, United States. 6. Dermatology, Complejo Asistencial Universitario de Salamanca, Salamanca, CL, Spain. Bioinformatics, Computational Biology, and Imaging