Authors - Mahzuzah Afrin, Rajasree Das Chaiti, Gazi Tahsina Sharmin Jahin, M. M. Musharaf Hussain, Mohammad Shamsul Arefin Abstract - Reliable identification of pneumonia from chest radiographs plays a central role in supporting clinical decision-making and patient management. Although deep learning models have shown favourable results for automated diagnosis, most existing studies rely on fully supervised training and mainly evaluate performance using accuracy or ROC-AUC metrics. Such evaluations may fail to capture clinical decision reliability, particularly in imbalanced medical datasets. In this work, we examine the effectiveness of self-supervised learning (SSL) for chest X-ray pneumonia classification through a controlled empirical study. A contrastive pretraining strategy is used to learn image representations from unlabeled chest X-rays, followed by supervised linear evaluation. The SSL-pretrained model is compared with a fully supervised model trained from scratch under identical experimental conditions. Our experiments reveal that the supervised baseline attains a slightly higher ROC-AUC; however, this improvement comes at the cost of increased false positive predictions, leading to lower overall accuracy. In contrast, the SSL-pretrained model exhibits a distinct prediction pattern. It achieves higher accuracy and notably improved precision and F1-score, indicating more balanced and reliable predictions. Precision– recall analysis further demonstrates the advantage of SSL in reducing false positive decisions. In addition, Grad-CAM visualizations suggest that the SSL-pretrained model focuses on clinically relevant lung regions. From a clinical decision-making perspective, these results suggest that self-supervised learning offers tangible advantages for chest X-ray analysis when prediction reliability is prioritized. This distinction is especially relevant in clinical settings.
Open Zoom