Adaptive sample selection with joint loss for medical image classification under label noise
Tongqing Xue, Aiping Qu, Han Hong · 2025
Deep Neural Networks (DNNs) have been extensively employed in the classification of medical images and achieve impressive performance. But the success of DNNs is dependent on the large amounts of correctly labelled images. However, noisy labels are often unavoidable in the real-world clinical scenarios, which significantly impact the performance of the model. In this manuscript, we introduce a new sample selection method which could select the clean samples adaptively without knowing the prior knowledge, such as label noise rates. We also integrated semi-supervised learning during sample selection to fully utilize the noised dataset. Specifically, we calculate batch statistics in each mini-batch and divide the samples into clean and noisy based on the statistics, then they are served as labelled and unlabeled in the semi-supervised manner. Furthermore, we use a joint loss to leverage useful information from unlabeled data along with a supervised loss, which strengthens the model's robustness. To evaluate the effectiveness of our method, we conduct sufficient experiments on a medical image dataset: Chaoyang. The results show that the proposed method could deal with the noisy labels in real-world scenarios.