Segmentasi Wilayah Abnormalitas Kanker Paru-Paru Menggunakan Residual Attention U-Net pada Citra Chest X-Ray

Abstract

Lung cancer is one of the leading causes of death worldwide, making early detection essential for effective treatment. Although chest X-ray screening is widely accessible, identifying subtle lung nodules remains difficult. This study develops a lung abnormality segmentation model based on the Residual Attention U-Net architecture and evaluates the effects of annotation techniques, training strategies, and input resolution on model performance using the JSRT dataset and local data from a hospital in Jember. Three annotation approaches applied. The model was trained using baseline training and a curriculum learning strategy based on subtlety levels using cropped image, followed by additional experiments using progressive input learning (PIL) and lung-field segmentation using full image. Annotation 3 combined with curriculum learning produced the best performance, achieving a Dice Score of 0.7362 and an IoU of 0.5825. The PIL strategy did not provide improvements because higher-resolution stages still needed to be resized due to computational limits, preventing the advantages of true high-resolution training. Lung-field segmentation increased performance by about 7% compared to training on full, uncropped images, indicating that restricting the model’s focus to the lung area can be beneficial, although it cannot be directly compared with the best cropped-input model due to differing input domains. The Residual Attention U-Net also outperformed baseline architectures, including U-Net, Residual U-Net, and Attention U-Net. When tested on local hospital data, the model experienced domain shift, but fine-tuning improved the Dice Score to 0.4559 and the sensitivity to 56%. These findings highlight the importance of realistic annotation, curriculum-based training, and architectural refinement in improving lung abnormality segmentation on chest X-ray images.