Boosting Few-Shot Learning with Disentangled Self-Supervised Learning and Meta-Learning for Medical Image Classification

TL;DR

This approach aims to enhance the model's generalization capabilities by exposing it to more challenging classification tasks during meta-training and evaluating it on easier tasks but holding greater clinical relevance during meta-testing, demonstrating the effectiveness and wide applicability of the proposed approach.

Abstract

Background and objective: Employing deep learning models in critical domains such as medical imaging poses challenges associated with the limited availability of training data. We present a strategy for improving the performance and generalization capabilities of models trained in low-data regimes. Methods: The proposed method starts with a pre-training phase, where features learned in a self-supervised learning setting are disentangled to improve the robustness of the representations for downstream tasks. We then introduce a meta-fine-tuning step, leveraging related classes between meta-training and meta-testing phases but varying the granularity level. This approach aims to enhance the model's generalization capabilities by exposing it to more challenging classification tasks during meta-training and evaluating it on easier tasks but holding greater clinical relevance during meta-testing. We demonstrate the effectiveness of the proposed approach through a series of experiments exploring several backbones, as well as diverse pre-training and fine-tuning schemes, on two distinct medical tasks, i.e., classification of prostate cancer aggressiveness from MRI data and classification of breast cancer malignity from microscopic images. Results: Our results indicate that the proposed approach consistently yields superior performance w.r.t. ablation experiments, maintaining competitiveness even when a distribution shift between training and evaluation data occurs. Conclusion: Extensive experiments demonstrate the effectiveness and wide applicability of the proposed approach. We hope that this work will add another solution to the arsenal of addressing learning issues in data-scarce imaging domains.

Figures (4)

• Figure 1: Illustration of the proposed approach. In the pre-training step, the feature backbone undergoes pre-training using the IP-IRM algorithm. At each iteration, the unlabeled dataset is divided into two subsets to maximize invariance between SSL losses. Subsequently, these subsets are employed to update the feature backbone parameters by minimizing the invariance between the SSL losses. The pre-trained backbone then undergoes meta-fine-tuning using the Meta DeepBDC algorithm. Meta-training episodes contain finer-grained classes while meta-testing coarser-grained ones belonging to the same source dataset. In both the meta-training and meta-testing phases a BDC matrix is computed for each support and query sample. Class prototypes are derived by averaging the BDC matrices of all support samples for that class. Classification is achieved by computing a similarity distribution of the query BDC matrix w.r.t. the class prototypes.
• Figure 2: Relationship between fine and coarse labels in (a) PI-CAI dataset and (b) BreakHis dataset.
• Figure 3: Results visual representation for (a) PI-CAI dataset and (b) BreakHis dataset. Each plot represents the results of a backbone in a 1-shot or 5-shot setting. We represented each fine-tuning scheme with a three-column group. In each group, the three colours indicate the three pre-training approaches.
• Figure 4: ROC curves for one random episode of meta-testing of each backbone pre-trained with SimCLR+IP-IRM and meta-fine-tuned in a 5-shot setting. Meta-training and meta-testing classes have the same source dataset (PI-CAI or BreakHis).