A Transformer-Residual Network-Integrated Deep Learning Model for Digital Pathology-Based Grading and Risk Stratification of Metabolic Dysfunction-Associated Steatotic Liver Disease-Related Liver Fibrosis.

INTRODUCTION

Metabolic dysfunction-associated steatotic liver disease (MASLD) affects up to 30% of the global population and drives progression to fibrosis, cirrhosis, and liver cancer, yet early-stage identification remains limited by subjective histopathology and low-sensitivity imaging. We aimed to develop and validate a deep learning model integrating Transformer and Residual Network architectures using digital histopathology images of human liver biopsies for MASLD-related fibrosis staging and identification of pathological phenotypes associated with advanced fibrosis, with interpretability analyses to support pathology-assisted decision-making.

METHODS

Four hundred sixty-seven Masson-stained liver biopsies from Duke (n = 338), Medizinische Hochschule Hannover (MHH) (n = 72), and Boehringer Ingelheim (BI)-int (n = 57) repositories were digitized at 20× resolution. Fibrosis was scored 0-4 per Younossi criteria. A four-encoder architecture combining 1 × 1 strided convolutions, multihead self-attention Transformer blocks, and Residual Network residual units was trained with global average pooling and linear classification. Performance metrics included accuracy, confusion matrices, continuous artificial intelligence (AI) scores, and gradient-weighted Class Activation Mapping visualization; analyses were conducted in Python/Jupyter with pandas, Matplotlib, and PyTorch.

RESULTS

Training accuracy reached 0.80 and validation 0.85 without overfitting. Stage-specific accuracy ranged 0.74-0.95, peaking at bridging fibrosis (stage 3). Gradient-weighted Class Activation Mapping heatmaps localized model attention to severe fibrotic septa, corroborating advanced fibrotic architectural changes. Continuous AI scores aligned tightly with discrete ground-truth labels, with boundary-adjacent misclassifications reflecting subtle histologic overlap.

DISCUSSION: The proposed model enables automated and interpretable assessment of MASLD-related fibrosis staging from digital liver biopsy histopathology and provides a quantitative tool for pathology phenotype-based risk stratification. Further clinical translation requires integration of clinical risk factors and validation in longitudinal outcome-based frameworks.