Classification of Atelectasis from Chest X-Ray Images Using Deep Learning Approaches

Abstract

Atelectasis

the partial or complete collapse of the lung

requires early diagnosis for effective treatment. This study investigates the effects of Region of Interest (ROI) cropping and transfer learning on the automatic classification of atelectasis from chest X-ray images. Four SqueezeNet-based models were developed using two image types (raw and cropped lung regions) and two training strategies (from-scratch and transfer learning). Models I and II were trained from scratch on raw and cropped images

respectively

while Models III and IV employed transfer learning on the same image sets. Performance was evaluated using standard classification metrics. Model IV

combining transfer learning and ROI-cropped images

achieved the best performance

with 98.43% accuracy

0.984 F1-score

and 0.999 AUC. It also reached 0.9861 precision

0.9828 sensitivity

and 0.9859 specificity. ROI-based cropping consistently improved classification performance. Combining transfer learning with ROI-focused preprocessing significantly enhances atelectasis detection and shows promise as a reliable clinical decision-support approach.