OMSF2: optimizing multi-scale feature fusion learning for pneumoconiosis staging diagnosis through data specificity augmentation

Abstract

Diagnosing pneumoconiosis is challenging because the lesions are not easily visible on chest X-rays, and the images often lack clear details. Existing deep detection models utilize Feature Pyramid Networks (FPNs) to identify objects at different scales. However, they struggle with insufficient perception of small targets and gradient inconsistency in medical image detection tasks, hindering the full utilization of multi-scale features. To address these issues, we propose an Optimized Multi-Scale Feature Fusion learning framework, OMSF2, which includes the following components: (1) Data specificity augmentation module is introduced to capture intrinsic data representations and introduce diversity by learning morphological variations and lesion locations. (2) Multi-scale feature learning module is utilized that refines micro-feature localization guided by heatmaps, enabling full extraction of multi-directional features of subtle diffuse targets. (3) Multi-scale feature fusion module is employed that facilitates the fusion of high-level and low-level features to better understand subtle differences between disease stages. Notably, this paper innovatively proposes a method for fine learning of low-resolution micro-features in pneumoconiosis, addressing the issue of maintaining cross-layer gradient consistency under multi-scale feature fusion. We established an enhanced pneumoconiosis X-ray dataset to optimize the lesion detection capability of the OMSF2 model. We also introduced an external dataset to evaluate other chest X-rays with complex lesions. On the AP-50 and R-50 evaluation metrics, OMSF2 improved by 3.25% and 3.31% on the internal dataset, and by 2.28% and 0.24% on the external dataset, respectively. Experimental results show that OMSF2 achieves significantly better performance than state-of-the-art baselines in medical image detection tasks.