Computer Aided Diagnosis CAD tool for lung and thoracic pathologies detection from chest X-ray images

Abstract

Detecting lung pathologies from chest radiographs effectively requires advanced feature representation. Imaging techniques such as X-rays provide essential, non-invasive insights into lung anomalies. Even the advances in deep learning have shown strong potential in improving lung pathology detection by autonomously identifying complex patterns in medical imaging, challenges remain, particularly due to the requirement for extensive labeled datasets and the sensitivity to data biases, which can affect model performance. This study introduces an advanced Deep Learning Based-Computer-Assisted Diagnosis tool for high-accuracy, automatic identification and classification of 14 distinct thoracic abnormalities from chest radiology images. Our methodology employs a two-stage approach, integrating the YOLOv10 object detection model with a pre-trained EfficientNetB0 architecture. This combination harnesses YOLOv10's robust detection capabilities and EfficientNetB0's efficient feature extraction, enabling precise analysis of lung and thoracic regions. Evaluated on a publicly available benchmark dataset, the system achieved a mean average precision equal to 0.488 at a 0.9, outperforming existing models. Additionally, it attained an overall precision of 0.864 across all pathologies, demonstrating its potential to enhance diagnostic decision-making as well as it could be adopted in clinical diagnosis routines to serve as a second opinion for radiologists.