Speaker
Description
Autoimmune Hepatitis (AIH) and Drug-Induced Liver Injury (DILI) cannot be reliably distinguished by pathologists from liver biopsy images alone. We investigate whether deep learning can identify image patterns that differentiate AIH from DILI in digitized Whole Slide Images (WSIs).
Patch-level features were extracted using the UNI foundation model and aggregated into slide-level predictions with an attention-based Multiple Instance Learning (MIL) framework.
The framework was trained on 108 WSIs (40 AIH and 68 DILI). The selected model achieved an AUROC of 0.812 on an internal test set and AUROCs of 0.96 and 0.85 on two independent external cohorts.
Ongoing work with expert pathologists focuses on interpreting highly attended regions and identifying morphological patterns associated with AIH and DILI.