Distribution-Aware Multi-Label FixMatch for Semi-Supervised Learning on CheXpert.

Sontje Ihler, Felix Kuhnke, Timo Kuhlgatz, Thomas Seel; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 2295-2304

Semi-supervised learning (SSL) has achieved remarkable success for multiclass classification in recent years yielding a promising solution for medical image classification where labeled data is scarce but unlabeled images are accessible. In the context of multi-label problems however SSL is still under-explored. In this work we adapt FixMatch to the multi-label scenario specifically focusing on CheXpert a multi-label chest X-ray classification dataset which is imbalanced and only partially labeled. Leveraging distribution alignment our proposed method ML-FixMatch+DA achieves solid performance gains in SSL tasks (AUC: +2.6%) and in a missing label scenario (AUC: +1.9%). In contrast to previous work we achieve a performance gain on CheXpert using FixMatch. We show that in contrast to multiclass FixMatch where distribution alignment is optional it is essential for multi-label FixMatch to handle class imbalance and generate reliable (positive and negative) pseudo-labels. Our pseudo-label selection is based on a single threshold for all classes and handles imbalance with no prior knowledge on label distributions. Our adaptation keeps the simplicity of the original multiclass FixMatch with no added hyperparameters (even for imbalanced data) and demonstrates the feasibility of simple SSL for multi-label problems filling a crucial gap in the literature.