Few-shot learning with task adaptation for multi-category gastrointestinal endoscopy classification

The prevalence of Gastrointestinal (GI) diseases exhibits a long-tailed distribution, resulting in the pervasive presence of highly imbalanced classes within real -world clinical GI endoscopy datasets. This poses a challenge for deep learning methods to train unbiased models capable of accurately classifying with limited labeled data, especially the minority classes. Moreover, GI endoscopy datasets collected from diverse clinical centers often exhibit distributional shifts and encompass distinct disease classes, thus significantly impeding the ability of deep learning methods to acquire models that can generalize well to unseen heterogeneous datasets. To address these challenges simultaneously, we approach the task of multi -category classification of GI endoscopic images on these real -world clinical datasets by formulating it as a few -shot learning (FSL) problem and propose the M etric -based F ew -shot learning with T ask A daptation (MFTA) multi -category classification approach to tackle this problem. Initially, we utilize self -supervised contrastive learning to obtain generalpurpose features that can serve as an optimal initial point for subsequent fine-tuning. Afterwards, we fine-tune the model by employing a metric based FSL method on the class imbalanced source dataset. To further improve generalization towards new classes, we integrate a novel episode augmentation mechanism. During transductive inference, we mitigate the domain gap between the source and unseen target datasets by employing an episode -wise adaptation layer. Additionally, we enhance class prediction through a prototype refinement strategy, thereby yielding superior performance gains. Extensive evaluations demonstrate the robust adaptability of our MFTA approach to novel classes with limited labeled data, as well as its generalization capability across diverse unseen datasets.