The problem of segmenting global glomerulosclerosis in gigapixel histopathological images: the borderless glomeruli

Accurately segmenting glomeruli in kidney whole slide images (WSIs) is essential for advancing automation in renal pathology but remains challenging in cases of global glomerulosclerosis, where Bowman’s capsule boundaries are often unclear. Conventional machine learning (ML) models perform well on normal glomeruli but struggle with sclerotic cases due to the lack of distinct structural cues. This study investigates the use of the foundation model segmentation generative pre-trained transformer (SegGPT) to address this limitation. We conducted experiments at both the patch and WSI levels on a private dataset to evaluate the performance of SegGPT foundation model against three non-foundation architectures, U-Net, U-Net3+, and SwinTransformer + U-Net, trained with and without fine-tuning. The study revealed high segmentation performance for normal glomeruli, with non-foundation models achieving mean Dice similarity coefficient (mDice) scores of up to 0.94. For segmental sclerosis, performance was moderate, with scores reaching up to 0.73. In contrast, the segmentation of globally sclerotic glomeruli proved substantially more challenging: Models trained only on normal samples yielded mDice scores below 0.03, and even with fine-tuning on mixed datasets, WSI-level performance remained limited (mDice < 0.16). With only few annotated examples, SegGPT demonstrated markedly superior performance in this scenario, achieving up to 0.43 at the WSI level and 0.74 at the patch level. However, its performance under idealized conditions also reveals limitations in clinical generalization. While conventional models perform well on normal and segmentally sclerotic glomeruli, their performance declines sharply in globally sclerotic cases, even with fine-tuning. SegGPT showed better generalization in these challenging scenarios, particularly at the patch level. However, its limited performance at the WSI level underscores the difficulty of translating patch-level accuracy to full-slide inference, where contextual ambiguity is greater. These results expose a persistent gap between controlled experimental setups and real-world conditions, reinforcing the need for more realistic evaluation protocols to advance clinical applicability.