Approach for Enhancing the Accuracy of Semantic Segmentation of Chest X-ray Images by Edge Detection and Deep Learning Integration

Lesia Mochurad ^*

• Lviv Polytechnic National University, Lviv, Ukraine

In modern medical diagnostics, the problem of accurately extracting structures with low contrast, a problem with existing segmentation methods, has been analyzed. To address this problem, the paper proposes a new approach that combines contour detection methods with a U-net deep learning model to enhance segmentation accuracy in chest X-ray images, focusing on lung, heart, and clavicle segmentation, which is essential for cardiothoracic disease diagnosis and reducing segmentation artifacts caused by overlapping structures. A method is developed to segment anatomical structures such as lungs, heart, and clavicles in chest X-ray images by integrating U-net with Sobel and Scharr filters for edge detection. The applied methods include stepwise processing, where contour detection improves the quality of the original data, and the U-net model performs accurate segmentation. The experimental results obtained showed a significant improvement in accuracy when using the Sobel filter compared to the baseline U-net model. For lung segmentation, an accuracy of 99.26 %, a Dice coefficient of 98.88 %, and a Jaccard index of 97.54 % were achieved, which exceeds the performance of previous methods. Similar results were obtained for heart segmentation – accuracy 99.47 %, Jaccard 94.14 %, and clavicles – accuracy 99.79 %, Jaccard 89.57 %. Thus, a method that integrates U-net with contour detection methods has been proposed, which provides a significant improvement in the quality of segmentation of the studied anatomical structures. In particular, the Sobel filter proved the most effective in improving segmentation in complex anatomical regions.