Abnormalities detection on chest radiograph with bounding box-based lungs extraction and object detection algorithm
Hai Thanh Nguyen,My N. Nguyen,Sang Chi Pham,Phuong Ha Dang Bui
DOI: https://doi.org/10.1007/s41870-023-01687-9
2024-02-21
International Journal of Information Technology
Abstract:Chest X-ray images are widely used to detect abnormalities in internal organs, including the heart, liver, lungs, and bones. However, interpreting these images necessitates specialized training and extensive professional experience. This study proposes a novel approach for detecting abnormalities in chest X-ray images. First, we employ a bounding box-based lung extraction approach using a Residual Network, ResNet-34, to extract the lung region and surrounding from the original chest X-ray image. Then, You Only Look Once, a state-of-the-art object detection model, is utilized to detect 14 common lung abnormalities, such as aortic enlargement, atelectasis, calcification, cardiomegaly, consolidation, interstitial lung disease (ILD), infiltration, lung opacity, nodule/mass, pleural effusion, pleural thickening, pneumothorax, pulmonary fibrosis, and other lesions, on new images with the extracted lung region. The experimental results on the VinDr-CXR dataset demonstrate the effectiveness of the proposed method in detecting and recognizing lung diseases. Furthermore, the proposed bounding box-based lung extraction method effectively reduces the size and eliminates redundant details of the chest X-ray images. Also, we obtain a slight improvement in detection performance (in AP@0.5 and AP@0.5:0.95 metrics) on the extracted lung areas compared to the original images.