Aprendizado auto-supervisionado para classificação e recuperação de imagens histopatológicas

Abstract

Histopathology analyzes tissues and lesions for disease diagnosis, but manual analysis is often laborious and error-prone. To improve accuracy and reduce the workload for pathologists, computer-aided diagnosis systems can be used, including disease classifiers and content-based image retrieval tools. However, to be used, these systems need to extract features from images, but labeling these images also requires a great deal of effort, and is often time-consuming and expensive. To mitigate these problems, approaches centered on self-supervised learning can be explored, taking advantage of the inherent structure of the data for learning. In this context, considering the outlined challenges, this study developed and experimentally evaluated self-supervised models applied to histopathological image classification and retrieval tasks. A comparative analysis was conducted using both pre-trained feature extractors trained on natural images—ResNet50, DINOv2 (Small and Base)—and stateof-the-art extractors specifically designed for histopathology, namely RetCCL and Phikon ViT-B. Furthermore, given the promising results obtained by the DINOv2 Small and Base models, fine-tuning was performed, leading to the development of DINOHist-S and DINOHist-B models to assess their effectiveness on histopathological images. To evaluate the potential of the models, extensive experiments were carried out with seven databases. The results in the lesion classification task demonstrated that, by fine-tuning the DINOv2 models with few iterations, it is possible to match or surpass the state-of-the-art extractors in the histopathological domain. For some scenarios, the evaluated models achieved efficiency above 99% for multiple measurements. Furthermore, it was observed that the state-of-the-art models were significantly inferior in kidney lesion databases, while fine-tuning allowed better results, even for small datasets. In the content-based image retrieval task, the results demonstrated that Phikon ViT-B outperformed the other models at multiple ranking levels, achieving MAP values above 90% in most databases and in all rankings evaluated. For most datasets, Phikon ViT-B was statistically superior, with the exception of the kidney lesion databases. For these, DINOHist-S and DINOHist-B presented the best results. In general, the results demonstrated that the adjustment of base models in specific domains can outperform specialized models trained on a large scale, requiring only a fraction of the resources and training time compared to state-of-the-art models. To extend this work, future research can be conducted, including the evaluation of new architectures, the use of other datasets and the analysis of different fine-tuning strategies, among others.