Self-Supervised Learning for Feature Extraction from Glomerular Images and Disease Classification with Minimal Annotations

What is it about?

Deep learning, a type of artificial intelligence, has shown great promise in analyzing kidney biopsy images. However, it typically requires large amounts of labeled data, which is time-consuming and expensive to create, especially in specialized fields like kidney pathology._x000D_ Our study explored a new approach called self-supervised learning, which can learn from images without needing labels. We used a method called DINO (self-distillation with no labels) on over 10,000 images of kidney glomeruli, which are crucial structures in the kidney._x000D_ We found that this approach could effectively identify different parts of the glomeruli without being explicitly taught what to look for. When we then used this pre-trained system to classify different kidney diseases and clinical features, it performed better than traditional methods, especially when we had limited labeled data available. To ensure our results were reliable, we also tested our method on two independent datasets from different hospitals. These tests confirmed that our approach worked well even on new, unseen data._x000D_ This research is important because it shows that we can create effective AI systems for analyzing kidney biopsies with less need for extensive labeling. This could make it easier and faster to develop AI tools for kidney pathology, potentially leading to quicker and more accurate diagnoses for patients with kidney diseases.

Featured Image