Classification and quantification of glomerular spike-like projections via deep residual multiple instance learning with multi-scale annotation

The primary pathological feature of Membranous Nephropathy (MN) is the presence of tiny spike-like projections on the glomerular basement membrane. Early detection and efficient treatment of spike-like projections are essential in halting disease progression. Renal pathology biopsy stands as the gold standard for diagnosing MN, and the accurate identification of glomerular spike-like projections plays a vital role in aiding diagnosis. Nevertheless, the tiny spike-like projection lesions and constraints in data quantity pose considerable challenges for supervised learning-based glomerular classification and quantification. We develop a Multi-Scale Annotation-based Multiple Instance Learning (MSA-MIL) model to address the issues. MSA-MIL utilizes image labels and box-level labels to jointly enhance the classification performance of the MIL model. Specifically, we first employ U-Net for glomerular image edge segmentation and subsequently train the MIL model on the dataset with image-level labels. Then, to overcome the limitations arising from the scarcity of positive instances and the relatively small size of spike-like projection features, we manually augment the number of instances with spike-like projections via using box-level annotation to further enhance the MIL model's classification performance. The designed MSA-MIL model enables the classification, visualization, and quantitative analysis of glomeruli with spike-like projections in renal pathology images. We validated and evaluated the designed MSA-MIL model. The model performed exceptionally well, achieving a high accuracy of 0.9847 and demonstrating a high recall rate, effectively preventing misdiagnosis. Additionally, we utilized heatmaps to visualize the locations of spike-like projections within glomeruli, enhancing the model's interpretability. Furthermore, through an analysis of the correlation between the stages of membranous nephropathy and the proportion of spike-like projections, we observed that as the disease advances, the proportion of spike-like projections increases. This finding serves to further validate the results obtained by the model. The MSA-MIL model is the first one specifically designed for classifying glomerular spike-like projections. It not only enhances classification performance but also proves to be more suitable for categorizing minute lesions compared to conventional Convolutional Neural Network (CNN) models. The visualization of glomerular lesions and the proportion of spike-like projections provide doctors with insights into the model's inference process, offering intuitive assistance for accurate diagnoses. This model brings significant hope for advancing research and diagnosis in the field of kidney diseases.