A LIGHTWEIGHT U-NET WITH SEPARABLE CONVOLUTIONS FOR EFFICIENT LUNG SEGMENTATION IN REAL-TIME MEDICAL IMAGING

Accurate and efficient segmentation of the lung regions is indispensable for detecting and managing pulmonary diseases, as it allows clinicians to identify abnormalities and plan effective intervention strategies. However, the high computational demands of many existing segmentation models pose a significant challenge, particularly for deployment in resource-constrained environments such as mobile, edge platforms, and point-of-care devices. Lung segmentation is further challenged by wide anatomical variability and imaging artifacts, which existing models often struggle to handle without access to large-scale hardware. This study addresses this limitation by introducing a lightweight U-Net architecture that integrates depthwise separable convolutions to reduce computational complexity while preserving segmentation accuracy. By replacing standard convolutional layers, the model achieves faster inference and significantly lower parameter counts, making it well-suited for IT applications in embedded systems and clinical informatics. The model was evaluated on the publicly available Pulmonary Chest X-Ray Defect Detection dataset from Kaggle, demonstrating its effectiveness in segmenting lung regions. The performance evaluation shows that our model delivers outstanding results, attaining a Dice score of 91.92%, a Jaccard index of 82.75%, precision of 92.64%, recall of 90.31%, and accuracy of 97.12% on the test dataset. These results highlight that the lightweight U-Net achieves state-of-the-art segmentation accuracy with significantly reduced computational overhead, making it ideal IT solution for real-time use in clinical workflows and deployment on limited-resource devices.