Abstract:Medical Image Analysis (MIA) is integral to healthcare, demanding advanced computational techniques for precise diagnostics and treatment planning. The demand for accurate and interpretable models is imperative in the ever-evolving healthcare landscape. This paper explores the potential of Self-Supervised Learning (SSL), transfer learning and domain adaptation methods in MIA. The study comprehensively reviews SSL-based computational techniques in the context of medical imaging, highlighting their merits and limitations. In an empirical investigation, this study examines the lack of interpretable and explainable component selection in existing SSL approaches for MIA. Unlike prior studies that randomly select SSL components based on their performance on natural images, this paper focuses on identifying components based on the quality of learned representations through various clustering evaluation metrics. Various SSL techniques and backbone combinations were rigorously assessed on diverse medical image datasets. The results of this experiment provided insights into the performance and behavior of SSL methods, paving the way for an explainable and interpretable component selection mechanism for artificial intelligence models in medical imaging. The empirical study reveals the superior performance of BYOL (Bootstrap Your Own Latent) with resnet as the backbone, as indicated by various clustering evaluation metrics such as Silhouette Coefficient (0.6), Davies-Bouldin Index (0.67), and Calinski-Harabasz Index (36.9). The study also emphasizes the benefits of transferring weights from a model trained on a similar dataset instead of a dataset from a different domain. Results indicate that the proposed mechanism expedited convergence, achieving 98.66% training accuracy and 92.48% testing accuracy in 23 epochs, requiring almost half the number of epochs for similar results with ImageNet weights. This research contributes to advancing the understanding of SSL in MIA, providing valuable insights for enhancing the reliability of artificial intelligence models in clinical applications.

Explainable rotation-invariant self-supervised representation learning

IRLSG: Invariant Representation Learning for Single-Domain Generalization in Medical Image Segmentation

Rotation-Oriented Collaborative Self-Supervised Learning for Retinal Disease Diagnosis

Any region can be perceived equally and effectively on rotation pretext task using full rotation and weighted-region mixture

Examining the quality of learned representations in self-supervised medical image analysis: a comprehensive review and empirical study

OPTiML: Dense Semantic Invariance Using Optimal Transport for Self-Supervised Medical Image Representation

Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging

A rotation meanout network with invariance for dermoscopy image classification and retrieval

Unsupervised Representation Learning by Predicting Image Rotations

Uncertainty aware and explainable diagnosis of retinal disease

Self-Supervised Representation Learning by Rotation Feature Decoupling

Rotation Has Two Sides: Evaluating Data Augmentation for Deep One-class Classification.

3D Self-Supervised Methods for Medical Imaging

Learning Self-Supervised Representations for Label Efficient Cross-Domain Knowledge Transfer on Diabetic Retinopathy Fundus Images

Rotation-Invariant Deep Embedding for Remote Sensing Images

Category-aware Feature Attribution for Self-Optimizing Medical Image Classification

Self-supervised learning for medical image analysis: Discriminative, restorative, or adversarial?

Effect of Rotation Angle in Self-Supervised Pre-training is Dataset-Dependent

Rotation and Translation Invariant Representation Learning with Implicit Neural Representations

Achieving Rotation Invariance in Convolution Operations: Shifting from Data-Driven to Mechanism-Assured

Self-Supervised Learning Featuring Small-Scale Image Dataset for Treatable Retinal Diseases Classification