Colo-ReID: Discriminative Representation Embedding with Meta-learning for Colonoscopic Polyp Re-Identification

TL;DR

Colo-ReID tackles colonoscopic polyp re-identification under few-shot, cross-domain conditions by introducing a meta-learning based embedding method and a dynamic Meta-Learning Regulation (MLR) to diversify meta-test features. The approach combines meta-training and meta-testing losses with standard ID and hard triplet losses, and replaces the last BN layer with a meta-learning regularization mechanism to generate domain-diverse features via Gaussian sampling and feature mixing. Empirical results on the Colo-Pair dataset show state-of-the-art mAP and Rank-5 performance, with notable gains over strong baselines and clear ablation-supported benefits from both Colo-ReID and MLR components. The work demonstrates the practicality of meta-learning for medical image retrieval with limited data and points to future directions in multi-domain, attention-based representation learning for robust polyp ReID in real-world settings.

Abstract

Colonoscopic Polyp Re-Identification aims to match the same polyp from a large gallery with images from different views taken using different cameras and plays an important role in the prevention and treatment of colorectal cancer. However, traditional methods for object ReID directly adopting CNN models trained on the ImageNet dataset usually produce unsatisfactory retrieval performance on colonoscopic datasets due to the large domain gap. Additionally, these methods neglect to explore the potential of self-discrepancy among intra-class or inter-class relations in the colonoscopic polyp dataset, which remains an open research problem in the medical community. To solve this dilemma, we propose a simple but effective training method named Colo-ReID, which can help our model learn more general and discriminative knowledge based on the meta-learning strategy in scenarios with fewer samples. Based on this, a dynamic Meta-Learning Regulation mechanism called MLR is introduced to further boost the performance of polyp re-identification. Our experimental results show that Colo-ReID consistently outperforms second-best method in terms of mAP performance by +2.3% on polyp re-identification task. Our source code is also publicly available at https://github.com/JeremyXSC/Colo-ReID.

Figures (8)

• Figure 1: Illustration of polyp re-identification task. Given the query video clip of colonoscopic polyp, colonoscopic video retrieval aims to accurately locate or find the similar clip which semantically corresponds to the given query.
• Figure 2: Description of the self-discrepancy of intra-class or inter-class relations for polyp dataset, which is caused the experience variations of the surgeon doctors, the variation in the shooting environment and the difference in patient condition, etc.
• Figure 3: Overview of the proposed Colo-ReID framework, which can randomly divide the sampled data into meta training set $\mathcal{D}_{mtr}$ and meta testing set $\mathcal{D}_{mte}$ respectively, and then enable our polyp ReID model to learn more robust and discriminative knowledge with the limited colonoscopy polyp matching data, which can significantly enhance the discriminative ability of our model.
• Figure 4: Schematic of the parameter update process using meta-training and meta-testing. Specifically, we randomly divide the sample set $\mathcal{D}$ into meta-training set $\mathcal{D}_{mtr}$ and the meta-testing set $\mathcal{D}_{mte}$. Additionally, we compute the meta-training loss $\mathcal{L}_{mtr}$ on the meta-training samples in the meta-training phase, and we also compute the meta-testing loss $\mathcal{L}_{mte}$ on the meta-testing samples in the meta-testing phase.
• Figure 5: Illustration of our Meta-Learning Regulation mechanism. Specifically, MLR generates more diverse meta-test features at the feature level, then replaces the last batch normalization layer of the model with a meta-learning regularization layer. Zoom in for the best view.