Efficient deep Learning object detection applied to the extraction of glomeruli
Over the past two decades, object detection has emerged as a critical task in computer vision research. It aims to quickly and accurately identify and locate a large number of objects of predefined categories in a given image. Many deep learning methods [1] have been introduced to address this complex task (Single Shot Multibox Detector, Yolo, Faster R-CNN, EfficientDet, Object Detection with Transformers, etc.).
Glomeruli (singular: glomerulus) are small, ball-shaped clusters of specialized capillaries found in the kidneys. They are a key component of the nephron, which is the functional unit of the kidney responsible for filtering waste products and excess substances from the blood to form urine. Accurate identification of glomeruli is essential not only to assess their density and distribution, but also to uncover any irregularities or potential pathology that could impact their performance. Consequently, the ability to detect glomeruli is a crucial phase in the complete analysis of kidney tissue.
For the experiments, two public datasets : https://www.kaggle.com/c/hubmap-kidney-segmentation/overview and https://aidpath.eu/ are available.
