Year: 2025
Type: conference-paper
Abstract: Glomeruli are central to the analysis of whole slide images (WSIs) in kidney biopsies, as they are affected by a wide range of lesions that reflect both causes and consequences of renal diseases. In automated WSI analysis using machine learning, glomeruli are typically the first regions to be segmented or detected, enabling subsequent diagnostic tasks. The Bowman’s capsule (BC) usually serves as the primary anatomical marker, delineating the glomerulus from the surrounding interstitium. While this boundary is preserved in normal and partially sclerotic glomeruli, globally sclerotic glomeruli often lose the BC, appearing visually borderless and posing a major challenge for automatic detection. In recent years, several studies have addressed glomerulus segmentation; however, few have focused on the specific challenge of globally sclerotic glomeruli, often analyzing them only in isolated per-cropped images. In this work, we present a comparative evaluation of four few-shot semantic segmentation (FSS) methods: DMACA, VAT, HSNet, and PMNet. These approaches aim to learn from only a few labeled examples, addressing the data scarcity of globally sclerotic glomeruli. These methods were applied to three classes of glomeruli: those with well-defined borders, partially borderless glomeruli, and globally sclerotic borderless glomeruli, using the Dice metric. Our results highlight the intrinsic difficulty of segmenting globally sclerotic glomeruli from WSIs, with a mean Dice score across all the evaluated methods of only 0.02 when evaluated at the wholeslide level. In contrast, per-crop evaluations yielded markedly higher performance, with mean Dice scores reaching 0.93 for globally sclerotic glomeruli.
Authors: Márcio dos Santos, Luiz Souza, Jefferson Fontinele, Marcelo Mendonça, Angelo Duarte, Washington L. C. dos-Santos, Luciano Oliveira
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7114766641