A Classification Method of the Crescent Status in IgA Nephropathy Based on Contrastive Learning

Immunoglobulin A nephropathy (IgAN) is a prevalent glomerulonephritis with an insidious onset. Early diagnosis and treatment significantly improve patient prognosis. Renal biopsy is the gold standard for diagnosis, but it is an invasive procedure with several limitations. Therefore, it is essential to explore a non-invasive diagnostic method. The crescent status is an independent predictor of IgAN progression, and subtle differences in CT scans make it difficult for CNNs to distinguish the progression. Contrastive learning enables CNNs to learn effective representations by comparing similarities and differences between samples, rather than relying on predefined labels. Consequently, we incorporate this approach into conventional CNNs to enhance their representational learning capabilities. In cases where the pathological characteristics of both kidneys exhibit similar features, the method utilizes the SimSiam contrastive learning method to compare images of two kidneys and combines it with supervised learning for end-to-end training. For supervised training, feature fusion is used to combine features from both kidneys into a more comprehensive representation, which avoid inconsistent results between the two kidneys, thereby accelerating model convergence. Experimental results demonstrate that CNN performance can be improved stably through applying the designed end-to-end training method of contrastive learning and feature fusion. Furthermore, we use a balancing factor to adjust the losses of the two training processes and explore its optimal range through experiments. Three balancing factor decay strategies are proposed and compared. Ultimately, the method proposed in this paper achieves an accuracy of 0.695 and an AUC of 0.699, demonstrating its potential to assist physicians in diagnosis.