A Multicenter Whole Slide Image Dataset for Classification and Out-of-Distribution Detection in Spindle Cell Cutaneous Neoplasms

Le résumé fourni par la source

Abstract Cutaneous spindle cell (CSC) neoplasms are a notoriously challenging diagnostic group within the spectrum of skin malignancies. While specialized CSC neoplasm datasets like AI4SKIN have enabled the development of AI whole-slide image (WSI) classification methods, these methods are largely constrained by closed-set designs. However, in real-world clinical practice, biopsies often contain secondary tumors or rare entities unknown to the developed AI classifiers. When faced with such Out-of-Distribution (OoD) cases, automated systems can produce highly confident but incorrect predictions, posing a significant risk to patient safety and hindering the reliable deployment of AI in routine pathology. To address this limitation, we present ASSIST, a multicenter dataset of 410 WSIs comprising both spindle-cell tumors and a diverse set of metastatic and rare lesions explicitly included as OoD samples. ASSIST expands AI4SKIN by increasing sample diversity, reinforcing underrepresented categories, and introducing clinically meaningful OoD cases, thereby enabling the development of more robust and safety-oriented computational pathology models. We describe the dataset acquisition pipeline, annotation structure, and technical validation using multiple instance learning (MIL) models combined with a suite of OoD detection methods. ASSIST provides an essential benchmark for advancing open-world pathology and supports future research in dependable AI systems for dermatopathology.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.