Enhancing Contrastive Training for Semi-Supervised Chest X-Ray Analysis Through Gaussian Mixture Models

Acquiring appropriate annotations for computer vision is a time-consuming and costly process, especially for medical imaging (i.e. chest radiographs) owed mostly to the necessity for trained radiologists. One solution to this problem can be seen in semi-supervised algorithms that use large amounts of unlabeled data to improve training convergence become essential. One particular solution, contrastive learning, which is a widely recognized method for efficient representation learning, is popular for learning descriptive representations of large datasets. By introducing a Gaussian mixture model, we introduce a trustworthiness factor for the inclusion of samples during contrastive learning. By using the resulting reliable samples, the model can learn more expressive representations and not succumb to semantic outliers. Our findings indicate that even without extensive pretraining, our contrastive approach improves the learning of representations of semi-supervised models, leading to a supervised performance of 81% AUC at just 20% of the labels.