FSD-Net: a fuzzy semi-supervised distillation network for noise-resistant classification of medical images
Article
Du, X, Shen, A, Wang, X et al. (2024). FSD-Net: a fuzzy semi-supervised distillation network for noise-resistant classification of medical images
. MULTIMEDIA TOOLS AND APPLICATIONS, 10.1007/s11042-024-18844-2
Du, X, Shen, A, Wang, X et al. (2024). FSD-Net: a fuzzy semi-supervised distillation network for noise-resistant classification of medical images
. MULTIMEDIA TOOLS AND APPLICATIONS, 10.1007/s11042-024-18844-2
Deep Learning (DL)-based models have been successfully applied for medical image classifications. However, the performance of traditional medical image classifiers is limited by insufficient training samples and inaccurate annotations due to labeling noise. Numerous methods were proposed to address the issues, to no avail. To fill this gap, we propose a Fuzzy Semi-supervised Distillation Network (FSD-Net) to improve the noise-resistant classification performance. Firstly, with the aid of semi-supervised training based on the pseudo-labels, the problem of insufficient training samples can be alleviated by using unlabeled data to a certain extent. Then, a Confidence Filtering model (CF-model) is designed to improve the quality of the pseudo-labels and filter the noisy labeling data by combining the soft labels generated from the classification model and the improved Soft Fuzzy C-Means (SFCM) model. Compared with the setting thresholds, the data with low confidence are reclassified as unlabeled to participate in training, while the labeled and unlabeled data (generating pseudo-labels) with high confidences are added into the mixed dataset. Finally, the mixed data are fed to train a multi-model online distillation module consisting of the SFCM and two other student models. The weighted average value of the three student models is used as teacher knowledge to train the students and obtain the final classification results. Experimental results on multiple medical datasets show that the proposed FSD-Net can solve the problems of insufficient labeled samples and labeling noise simultaneously to some extent.
