Predicting cSCC grade using RACR-MIL: A comparative study with dermatopathologists

Summary: Abstract Body: Cutaneous squamous cell carcinoma(cSCC) is common, yet current staging systems often fail to predict poor outcomes; prior studies have shown up to 46% rate of under-staging of high risk cases. Tumor differentiation is crucial for staging and pathologists use variable approaches to assess it. More consistent and accurate identification of tumor grade is needed to recognize tumors at risk for poor outcomes. Eighteen histopathologically challenging cases were annotated and interpreted by 5 dermatopathologists and grouped by consensus diagnosis. We used a rank-aware contextual reasoning-based machine learning model(RACR-MIL) to predict cSCC grade using whole-slide images and examined its impact upon inter-rater reliability and consensus diagnosis. Majority consensus (agreement among ≥3 pathologists) occurred in 12 cases (66.7%). Minority consensus (agreement among a significant minority or combined adjacent groupings) occurred in 6 cases (33.3%). The model’s prediction achieved an overall agreement of 77.8% with consensus, aligning with the majority consensus in 10/12 cases (83.3%) and the minority consensus in 4/6 (66.7%) cases. RACR-MIL shows promising results with up to 50% higher alignment with pathologist’s annotations compared to baselines, capturing subtle features such as nuclear atypia in inflammatory cells and infiltrative patterns at the tumor edge.These results highlight the variation inherent in cases and the challenge of assigning a grade at the whole-slide level. Future directions for reporting grading as continuous variables to reflect tumor heterogeneity are warranted. Anirudh Choudhary¹, Alyssa L. Stockard², Zachary Leibovit-Reiben², Angelina Hwang², Jacob Kechter², Krishnakant Saboo¹, Nneka Comfere⁴, Steven Nelson², Emma Johnson⁴, Leah Swanson², Olayemi Sokumbi⁵, Jason Sluzevich⁵, Collin Costello², Paul Arnold⁶, Javier Canueto³, 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, Complejo Asistencial Universitario de Salamanca, Salamanca, CL, Spain. 4. Dermatology, Mayo Clinic Minnesota, Rochester, MN, United States. 5. Dermatology, Mayo Clinic in Florida, Jacksonville, FL, United States. 6. Carle Illinois College of Medicine, Urbana, IL, United States. Bioinformatics, Computational Biology, and Imaging