Self-supervised Mean Teacher for Semi-supervised Chest X-Ray Classification

Abstract

The training of deep learning models generally requires a large amount of annotated data for effective convergence and generalisation. However, obtaining high-quality annotations is a laboursome and expensive process due to the need of expert radiologists for the labelling task. The study of semi-supervised learning in medical image analysis is then of crucial importance given that it is much less expensive to obtain unlabelled images than to acquire images labelled by expert radiologists.Essentially, semi-supervised methods leverage large sets of unlabelled data to enable better training convergence and generalisation than if we use only the small set of labelled images.In this paper, we propose the Self-supervised Mean Teacher for Semi-supervised (S 2 MTS 2 ) learning that combines self-supervised mean-teacher pre-training with semi-supervised finetuning. The main innovation of S 2 MTS 2 is the self-supervised mean-teacher pretraining based on the joint contrastive learning, which uses an infinite number of pairs of positive query and key features to improve the mean-teacher representation. The model is then fine-tuned using the exponential moving average teacher framework trained with semi-supervised learning.We validate S 2 MTS 2 on the thorax disease multi-label classification problem from the dataset Chest X-ray14, where we show that it outperforms the previous SOTA semi-supervised learning methods by a large margin. Our code is available at https://github.com/ fengbeiliu/semi-chest.

Augmentation-Consistent Clustering Network for Diabetic Retinopathy Grading with Fewer Annotations

“…Besides, we choose two existing semisupervised medical image classification methods [ 30 , 41 ] to compare with our ACCN model. S 2 MTS 2 [ 30 ] combines self-supervised mean-teacher pretraining with a semisupervised fine-tuning method to solve the multilabel chest X-ray classification; SRC-MT [ 41 ] proposes a sample relation data consistency paradigm to effectively extract unlabeled data by modeling the relationship information among different medical image samples. To compare the ACCN with them, we implement their public available code on the Messidor dataset with the same settings.…”

“…Pang et al [ 29 ] developed a radionics model based on a semisupervised GAN method to perform data augmentation in breast ultrasound images. Liu et al [ 30 ] proposed a self-supervised mean teacher for chest X-ray classification that combines self-supervised mean-teacher pretraining with semisupervised fine-tuning. Bakalo et al [ 31 ] designed a deep learning architecture for multiclass classification and localization of abnormalities in medical imaging illustrated through experiments on mammograms.…”

Self-Supervised Pretraining Enables High-Performance Chest X-Ray Interpretation Across Clinical Distributions

“…To evaluate whether self-supervised CXR pretraining outperforms ImageNet pretraining, we selected seven SSL methods, three that only use CXR images (MedAug 16 , S2MTS2 17 , MoCo-CXR 18 ) and four that use both chest X-ray images and corresponding radiology reports (CXR-RePaiR-CLIP 19 , ConVIRT 20 ,REFERS 21 , GLoRIA 22 ). We give further details on these learning algorithms in the Methods section.…”

A Review of Predictive and Contrastive Self-supervised Learning for Medical Images

“…Teacher-student learning for medical applications -Li et al [156] designed a new SSL approach based on the teacher-student architecture to learn distinguishing representations from gastric X-ray images for a downstream task, gastritis detection. One of the student-teacher frameworks, Mean Teacher in [157], was integrated by Liu et al [158] in the pretraining process for semisupervised fine-tuning for thorax disease multilabel classification. Park et al [159] used information distillation between teacher and student frameworks and the vision transformer model for chest X-ray diagnosis, including tuberculosis, pneumothorax, and COVID-19.…”