Localization and classification of abnormalities on chest X-ray images using a Mamba-YOLOvX model

Abstract

Chest X-rays (CXR) are critical diagnostic tools for detecting thoracic abnormalities. However, challenges such as overlapping anatomical structures, class imbalance, and dataset heterogeneity hinder accurate interpretation and limit model generalizability. To address these issues, a Mamba-YOLOvX model is presented in this study. It was aimed to integrate global and local lesion information to improve the detection and localization of thoracic abnormalities. The model incorporates novel architectural improvements, including combined spatial and channel attention mechanisms and selective scanning blocks, to capture fine-grained features and enhance multi-scale detection. In addition, a projection-based data augmentation strategy, leveraging rib segmentation and keypoint alignment was developed to improve the anatomical consistency and the intensity normalization across datasets. Extensive experiments were conducted on three large-scale datasets (VinDr-CXR, ChestX-ray8, and CXR-AL14), achieving state-of-the-art performance in detecting abnormalities of varying sizes. The proposed method reached an average precision at 50 % intersection over union of 0.366, 0.153, and 0.615 on the VinDr-CXR, ChestX-ray8, and CXR-AL14 datasets, respectively. Results demonstrated significant improvements in precision, recall, and mean average precision, particularly for small lesions. Cross-dataset validation confirmed the model’s robustness and generalizability. This study highlights the potential of integrating advanced deep learning techniques with domain-specific augmentations to enhance clinical decision support systems for thoracic disease detection. By addressing critical challenges such as class imbalance, annotation inconsistencies, and scale variations, the enhanced Mamba-YOLOvX model is shown as a scalable, accurate, and generalizable solution for automated CXR analysis.