Classification of atelectasis from chest X-ray images using deep learning approaches

In this study, deep learning methods were employed to automatically classify atelectasis disease from chest X-ray images. Atelectasis, characterized by the partial or complete collapse of the lung, is a serious respiratory condition where early diagnosis plays a critical role in the treatment process. The study aimed to investigate the impact of different learning strategies and image preprocessing techniques on classification performance by developing four distinct models. The models were differentiated based on two main variables: (1) whether raw images or cropped images containing only the lung region were used, and (2) whether the model was trained from scratch or utilized a pre-trained (transfer learning) architecture. Two SqueezeNet-based models were trained from scratch (Model I and Model II), while two models employed transfer learning using a pre-trained SqueezeNet architecture (Model III and Model IV). Models I and III were trained on raw images, whereas Models II and IV were trained on cropped images containing only the lung region. Experimental results revealed that the highest performance was achieved by Model IV, with an accuracy of 98.43%, an F1 score of 0.984, and an AUC of 0.999—outperforming many existing methods in the literature. Additionally, cropping the images improved overall performance in both from-scratch and transfer learning-based models. The findings demonstrate that focusing on the region of interest and utilizing transfer learning provide significant advantages in medical image classification. This study shows that deep learning-based systems can serve as reliable decision support tools for the early detection of critical diseases such as atelectasis.