Um estudo abrangente sobre segmentação de glomérulos com dados de treinamento limitados em imagens histopatológicas de alta resolução

Abstract

The growing availability of scanned whole slide images (WSIs) has expanded digital pathology, enabling medical decision-making and computational analysis directly from high-resolution images. Kidney disease diagnoses using WSIs rely on the analysis of specific tissue structures, and automatic analysis depends on accurately segmenting key components such as glomeruli, tubules, interstitium, and vessels. This thesis focuses on glomeruli, which are essential in assessing WSIs after kidney biopsies. These structures are impacted by lesions related to various diseases. In machine learning-based WSI analysis, glomeruli are often the first regions segmented to guide subsequent tasks. The Bowman’s capsule (BC) is crucial, marking the boundary between glomerular components and surrounding interstitial tissue. This work proposes two studies aimed at addressing the segmentation of glomeruli in high-resolution kidney histopathological images. In the first study, we investigate the feasibility of segmenting glomeruli in human WSIs using deep-learning models trained exclusively on mouse data. Mice and humans share several biological similarities, including genetic, physiological, and structural characteristics, making mice a common model for studying human diseases. While this cross-species knowledge transfer is well-established in medicine, it remains underexplored in computational pathology, where WSIs serve as primary research objects. To address this gap, we evaluated five semantic segmentation models: U-Net, U-Net 3+, Res-U-Net, DeepLabV3+, and MA-Net, using datasets consisting of 18 mouse WSIs and 42 human WSIs. Among these, U-Net 3+ delivered the best performance in intra-dataset evaluation, achieving an average DICE score of 0.930 on HE-stained mouse images. On human data, U-Net 3+ also excelled, attaining DICE scores of 0.772, 0.824, and 0.791 on HE, PAS, and PAMS stains, respectively. Moreover, U-Net 3+ proved promising generalization when trained solely on mouse data and tested across the entire human dataset, achieving a DICE score of 0.798 on HE-stained images. However, while these models performed well on images within the same staining technique, their performance declined when applied across different stains, highlighting a limitation in cross-stain generalization. The second study focuses on the segmentation challenges posed by borderless glomeruli affected by global sclerosis. We developed an automated framework for patch cropping and stitching, eliminating manual intervention to streamline the segmentation process. Our experiments show that while standard segmentation models can achieve state-of-the-art results for normal and partially sclerotic glomeruli, their performance deteriorates significantly for globally sclerotic glomeruli. Notably, segmentation accuracy for these cases was highly dependent on the staining type and generally remained poor across models. We compared non-foundation models (U-Net, U-Net 3+, and SwinTransformer + U-Net) with and without fine-tuning against the SegGPT foundation model. Non-foundation models, trained exclusively on normal glomeruli with HE, PAS, and PAMS stains, achieved high performance on normal glomeruli (mDice > 0.92) and moderate performance on partially sclerotic glomeruli (mDice > 0.72). However, their performance dropped sharply to mDice > 0.02 for globally sclerotic glomeruli, with minimal improvements even after fine-tuning. In contrast, SegGPT demonstrated substantial improvement, achieving a significantly higher mDice score (> 0.37) for globally sclerotic glomeruli by leveraging only a few query samples. This result highlights the potential of foundation models in addressing segmentation challenges for glomeruli affected by severe lesions. In summary, the studies presented in this thesis represent a significant step forward in the segmentation of glomeruli in WSIs. Our findings offer a comprehensive analysis of glomerulus segmentation with limited training data, demonstrating the potential of mouse-to-human transfer learning, as well as the use of foundation models to improve segmentation accuracy for glomeruli affected by sclerosis.