Accurate histological differentiation of hepatocellular carcinoma (HCC) is essential for informing clinical decision-making, yet conventional grading remains labor-intensive and susceptible to inter-observer variability among pathologists.¹ While artificial intelligence (AI) has demonstrated potential in computational pathology, the lack of systematic comparison of AI frameworks for HCC grading limits their adoption in clinical practice.¹ At the AASLD Annual Meeting 2025, Professor Seto, Wai-Kay from the University of Hong Kong, presented a novel AI-based framework utilizing multiple instance learning (MIL) on digitized histopathology slides, offering a potential solution to standardized and efficient HCC grading.¹

Digital histology, which converts glass or paraffin slides into high-resolution whole slide images (WSIs) for virtual viewing, enables remote collaboration, objective analysis, and efficient storage and retrieval.¹ Utilizing these WSIs, the study was conducted to evaluate and compare AI frameworks based on MIL strategies for aggregating patch-level features to optimize automated, slide-level histological grading of HCC.¹

This study analyzed WSIs derived from 54 surgically resected HCC patients, comprising a total of 392 slides that included tumors and non-tumor tissue, graded according to the American Joint Committee on Cancer (AJCC) 8th edition criteria.¹ Tumor differentiation was determined by an experienced liver histopathologist and categorized as well- or moderately differentiated (G1/G2) or poorly differentiated (G3).¹ WSIs were segmented into 1024×1024-pixel patches, with patches containing tumor tissue labeled according to histological differentiation.¹

Following patch extraction,  each patch was processed using a pretrained foundation model UNI, a large vision transformer trained on over 100,000 hematoxylin-and-eosin (H&E)-stained WSIs across multiple organ systems, to generate patch-level feature embeddings.¹ Uniform manifold approximation and projection (UMAP) was then applied to patch-level embeddings for visualization of feature distributions and qualitative assessment of histological patterns.¹ These embeddings were subsequently aggregated at the slide level using MIL to enable slide-level grading.¹

MIL requires only a slide-level label (positive/negative or G1-G3) and automatically identifies the most discriminative regions contributing to the final slide level prediction.¹ Three MIL-based aggregation frameworks were evaluated for slide-level grading in the study: logistic regression as a baseline method, attention-based MIL, and transformer-based MIL.¹ Model performance was assessed using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).¹

Visualization using UMAP demonstrated that the UNI feature extractor effectively identified meaningful histological patterns.¹ Morphologically similar tissue patches clustered closely, with clear separation between non-tumor tissue and HCC tumor regions.¹ This finding supports the suitability of the pretrained foundation model for downstream histopathological analysis.¹

Across all evaluated metrics, both advanced MIL approaches outperformed the baseline logistic regression model.¹ The transformer-based MIL model achieved the highest overall performance, with an AUC of 0.888 (95% CI: 0.706-0.998), sensitivity of 0.892 (95% CI: 0.784-0.976), and specificity of 0.957 (0.913-0.988) for slide-level differentiation grading.¹ The model achieved a PPV of 0.891 (95% CI: 0.780-0.974) and an outstanding NPV of 0.957 (95% CI: 0.915-0.987), underscoring its robustness in accurately excluding negative cases.¹ The superior performance of transformer-based MIL was attributed to its ability to model contextual relationships among patches, allowing the algorithm to integrate spatial and co-occurring patterns across the entire slide rather than relying on isolated features.¹

Differentiation-specific analyses further highlighted the robustness of the approach.¹ The transformer-based MIL model achieved an AUC of 0.926 (95% CI: 0.831-0.985) for non-tumor tissue, 0.881 (95% CI: 0.784-0.949) for well- and moderately differentiated tumors (G1/G2), and 0.924 (95% CI: 0.781-0.995) for poorly differentiated tumors (G3).¹ Notably, performance remained strong even for G3 tumors, which are often more heterogeneous and challenging to classify.¹ High negative predictive values across differentiation categories suggest reliable exclusion of incorrect grades, supporting potential clinical utility in assisting diagnostic workflows.¹

In conclusion, MIL frameworks surpassed traditional probability aggregation, and the transformer model’s ability to process long patch sequences further enhanced diagnostic accuracy.¹ Integrating MIL with pretrained foundation models significantly improves slide-level HCC differentiation grading by effectively capturing histological patterns across WSIs.¹ The ability of transformer-based MIL to model contextual relationships further underscores its potential for complex computational pathology tasks.¹ With future validation in larger, multi-center cohorts, this AI-driven approach may represent an important cornerstone toward standardized, scalable histological assessment in HCC.¹