Explainable self-supervised learning for medical image diagnosis based on DINO V2 model and semantic search.
Authors
Affiliations (3)
Affiliations (3)
- Machine Learning and Information Retrieval Department, Faculty of Artificial Intelligence, Kaferelshikh University, Kaferelshikh, 33511, Egypt.
- Biological Artificial Intelligence Program, Faculty of Artificial Intelligence, Kafer Elsheikh University, Kaferelshikh, Egypt.
- Machine Learning and Information Retrieval Department, Faculty of Artificial Intelligence, Kaferelshikh University, Kaferelshikh, 33511, Egypt. [email protected].
Abstract
Medical images have become indispensable for decision-making and significantly affect treatment planning. However, increasing medical imaging has widened the gap between medical images and available radiologists, leading to delays and diagnosis errors. Recent studies highlight the potential of deep learning (DL) in medical image diagnosis. However, their reliance on labelled data limits their applicability in various clinical settings. As a result, recent studies explore the role of self-supervised learning to overcome these challenges. Our study aims to address these challenges by examining the performance of self-supervised learning (SSL) in diverse medical image datasets and comparing it with traditional pre-trained supervised learning models. Unlike prior SSL methods that focus solely on classification, our framework leverages DINOv2's embeddings to enable semantic search in medical databases (via Qdrant), allowing clinicians to retrieve similar cases efficiently. This addresses a critical gap in clinical workflows where rapid case The results affirmed SSL's ability, especially DINO v2, to overcome the challenge associated with labelling data and provide an accurate diagnosis superior to traditional SL. DINO V2 provides 100%, 99%, 99%, 100 and 95% for classification accuracy of Lung cancer, brain tumour, leukaemia and Eye Retina Disease datasets, respectively. While existing SSL models (e.g., BYOL, SimCLR) lack interpretability, we uniquely combine DINOv2 with ViT-CX, a causal explanation method tailored for transformers. This provides clinically actionable heatmaps, revealing how the model localizes tumors/cellular patternsa feature absent in prior SSL medical imaging studies Furthermore, our research explores the impact of semantic search in the medical images domain and how it can revolutionize the querying process and provide semantic results alongside SSL and the Qudra Net dataset utilized to save the embedding of the developed model after the training process. Cosine similarity measures the distance between the image query and stored information in the embedding using cosine similarity. Our study aims to enhance the efficiency and accuracy of medical image analysis, ultimately improving the decision-making process.