Semi-supervised Medical Image Classification with Relation-driven Self-ensembling Model (2005.07377v1)

Abstract: Training deep neural networks usually requires a large amount of labeled data to obtain good performance. However, in medical image analysis, obtaining high-quality labels for the data is laborious and expensive, as accurately annotating medical images demands expertise knowledge of the clinicians. In this paper, we present a novel relation-driven semi-supervised framework for medical image classification. It is a consistency-based method which exploits the unlabeled data by encouraging the prediction consistency of given input under perturbations, and leverages a self-ensembling model to produce high-quality consistency targets for the unlabeled data. Considering that human diagnosis often refers to previous analogous cases to make reliable decisions, we introduce a novel sample relation consistency (SRC) paradigm to effectively exploit unlabeled data by modeling the relationship information among different samples. Superior to existing consistency-based methods which simply enforce consistency of individual predictions, our framework explicitly enforces the consistency of semantic relation among different samples under perturbations, encouraging the model to explore extra semantic information from unlabeled data. We have conducted extensive experiments to evaluate our method on two public benchmark medical image classification datasets, i.e.,skin lesion diagnosis with ISIC 2018 challenge and thorax disease classification with ChestX-ray14. Our method outperforms many state-of-the-art semi-supervised learning methods on both single-label and multi-label image classification scenarios.

Semi-Supervised Medical Image Classification with Relation-driven Self-ensembling Model

The paper addresses the challenge of medical image classification within a semi-supervised learning framework, focusing specifically on the effective exploitation of unlabeled data. Semi-supervised learning (SSL) has long been recognized as a beneficial approach within domains where labeled data collection is labor-intensive and expensive. In medical imaging, where professional expertise is necessary for accurate annotation, the ability to utilize large-scale unlabeled datasets affords significant practical advantages.

The authors introduce a novel framework that leverages relation-driven self-ensembling models to improve classification performance, notably in scenarios with sparse labeling. The framework builds upon the Mean Teacher (MT) model, renowned for yielding high-quality consistency targets through ensemble averaging. The central innovation involves a Sample Relation Consistency (SRC) paradigm, which models and preserves intrinsic relationships among input samples across perturbations.

The SRC paradigm represents a departure from conventional prediction consistency approaches, which maintain individual-level consistency without considering inter-sample relationships. By encouraging semantic relation matrices derived from high-level features to remain stable across perturbations, the framework can extract additional semantic information from unlabeled data, enhancing overall performance.

Experiments conducted on skin lesion and chest X-ray datasets present compelling evidence for the framework’s effectiveness. The framework demonstrates superior performance in both single-label and multi-label classifications, achieving higher AUCs and sensitivity scores than several state-of-the-art SSL methods, including self-training, Generative Adversarial Networks, and consistency-based methods such as Temporal Ensembling.

The results underscore the value of incorporating relational information, a strategy reflecting the diagnostic processes of clinicians—who often reference analogous cases to inform their decisions. By aligning model predictions with inherent data structures, the framework boosts the robustness and discriminability of representations.

The paper’s contributions are marked by three key advances:

1. Relation-driven SSL Framework: A novel framework that maintains effectiveness across various classification contexts, highlighting its adaptability to different medical imaging scenarios.
2. Sample Relation Consistency Paradigm: A methodological innovation facilitating the extraction of deeper semantic insights from unlabeled data through structured inter-sample relations.
3. Empirical Validation: Extensive testing confirms the framework’s advantages over existing SSL models, demonstrating consistent performance increases across different labeled data proportions.

Future research directions may delve into alternative approaches for estimating sample relations, further enhancing the performance and applicability of SSL across wider medical imaging applications. Integrating automatic augmentation techniques or more advanced network architectures may also hold promise for increasing classification efficacy under limited supervision. Additionally, exploring the paradigm’s potential as a data augmentation technique in fully supervised contexts could present intriguing avenues for enhancing model generalization.

In summary, the paper contributes a robust, innovative strategy for semi-supervised medical image classification, aligning model learning with clinical diagnostic practices through relation-driven techniques. It paves the way for reducing the reliance on labeled datasets, ultimately supporting more efficient and scalable medical image analysis solutions.