Quo Vadis Pathology? Advancing Glomerular Lesion Classification with Foundation Models

• David Lima UFBA
• Grinaldo Oliveira UFBA
• Ângelo Duarte UEFS
• Washington Santos FIOCRUZ
• Luciano Oliveira UFBA / FIOCRUZ

Resumo

Computational pathology is undergoing a significant transformation with the emergence of foundation models (FMs). These models leverage self-supervised learning on extensive histopathological datasets with the aim of extracting robust feature representations. FMs hold potential to automate advanced diagnostic pipelines, encompassing segmentation, classification, and biomarker discovery. This study evaluates the effectiveness of embeddings from four SOTA FMs (UNI, UNI2, Phikon, and Phikon2) for one-versus-all glomerular lesion classification. We propose here a comparative framework in which a multi-layer perceptron (MLP) and a support vector machine (SVM) - each trained exclusively on FM-derived embeddings - are benchmarked against EfficientNet, a fully supervised end-to-end image classifier. By varying the number of cross-validation folds (from k=2, representing minimal training data, to k=5, representing maximal training data), on a proprietary histopathology dataset, we assess classifier robustness under differing data regimes. Our results demonstrate that, even without any FM fine-tuning, the UNI/SVM pipeline outperforms the EfficientNet by 3.4 percentage points in average F 1-score, considering all values of k.

Palavras-chave: Support vector machines, Frequency modulation, Foundation models, Histopathology, Pipelines, Training data, Self-supervised learning, Feature extraction, Robustness, Lesions