Progressive Multi-Scale Learning Framework Improves Multi-Disease Detection in Chest X-Rays

A lightweight deep-learning framework termed PMSL-YOLO has demonstrated improved multi-disease detection performance in chest radiographs while maintaining the low computational overhead of its underlying YOLOv8n architecture, according to an original research study published in Frontiers in Medicine under the section on Medical Imaging and Nuclear Medicine. Developed by researchers Shanwei Chen, Quanhong Wang, Lihong Zhu, and Fengmei Su—affiliated with Baoji University of Arts and Sciences, the 96607 Hospital of the PLA, and the Yibin Engineering Technology Research Center for Intelligent Fluid Measurement and Control—the framework implements a structured Progressive Multi-Scale Learning strategy. Across an independent evaluation using the VinBigData dataset, the progressive detection model achieved higher diagnostic metrics than a conventional baseline detector, offering a potential solution for resource-constrained clinical environments where computational capacity is restricted.

Chest X-ray imaging serves as a primary diagnostic imaging modality for population-level thoracic disease screening, yet manual interpretation of complex multi-disease cases remains inherently challenging in routine clinical practice. Real-world diagnostic workflows are frequently constrained by heavy radiologist workloads and marked variability in diagnostic experience among reviewing clinicians, operational factors that heighten the clinical risk of missed abnormalities.

Although deep learning-based detectors have achieved promising detection performance in thoracic imaging evaluations, many high-accuracy models remain computationally intensive to execute. This substantial computational burden creates significant practical barriers to deployment within resource-constrained clinical environments that lack expensive, high-performance computing hardware. Furthermore, lightweight neural network detectors designed for rapid execution face pronounced technical challenges in learning effective multi-scale lesion representations under conventional fixed-resolution training paradigms.

To address the limitations inherent to conventional fixed-resolution training strategies, the investigators developed PMSL-YOLO, a lightweight framework constructed upon the compact YOLOv8n architecture and driven by Progressive Multi-Scale Learning. Rather than maintaining a static image resolution throughout the training cycle, the PMSL approach introduces a structured progressive learning paradigm that gradually guides the detector from global anatomical representation learning to fine-grained lesion localization through multi-resolution adaptation. The framework strategically incorporates progressive resolution transitions and staged optimization strategies to systematically improve feature representation capability. This staged training structure enables the lightweight network to incrementally refine its feature extraction mechanisms across different spatial scales without altering the fundamental computational efficiency of the underlying model.

In addition to the multi-resolution training schedule, the research team integrated specialized data augmentation strategies tailored specifically to the challenges of chest X-ray interpretation. These chest X-ray-specific augmentation methods were explicitly employed to enhance the detection of subtle and small lesions that might otherwise escape automated identification. The training pipeline incorporated optimized Mosaic augmentation, MixUp image blending, hue-saturation-value brightness adjustment, and moderate geometric transformations.

When tested on an independent test set of chest radiographs within the VinBigData experimental protocol, PMSL-YOLO configured with a progressive multi-scale resolution sequence of 512, 512, 640, and 1024 pixels demonstrated substantial performance gains over the baseline detector. The progressive PMSL-YOLO framework achieved a precision of 94.1 percent, a recall of 92.4 percent, a mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5) of 93.3 percent, and an mAP@0.5:0.95 of 70.1 percent. In direct comparison, the standard 1024 by 1024 baseline YOLOv8n model trained under a fixed-resolution paradigm achieved a precision of 89.6 percent, a recall of 80.9 percent, an mAP@0.5 of 86.7 percent, and an mAP@0.5:0.95 of 68.4 percent, demonstrating that the progressive resolution transition strategy notably improved both precision and sensitivity.

The authors concluded that PMSL-YOLO establishes an effective lightweight progressive training strategy for multi-disease chest X-ray detection, achieving improved detection performance while preserving the lightweight computational characteristics of YOLOv8n. The independent test results provide clear empirical evidence of observed performance gains on unseen data not utilized during model development within the VinBigData dataset and experimental protocol. However, the researchers emphasized that crucial clinical questions regarding generalizability remain unresolved.

Subject of Research: Lightweight progressive multi-scale deep learning for multi-disease chest X-ray detection

Article Title: Progressive multi-scale learning for lightweight YOLO-based multi-disease detection in chest X-ray

Article References: Chen, S., Wang, Q., Zhu, L., & Su, F. (2026). Progressive multi-scale learning for lightweight YOLO-based multi-disease detection in chest X-ray. Frontiers in Medicine, 13, Article 1901057. https://doi.org/10.3389/fmed.2026.1901057

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1901057

Keywords: chest X-ray, multi-disease detection, PMSL-YOLO, YOLOv8n, progressive multi-scale learning, medical imaging, computer-aided diagnosis, VinBigData dataset, Progressive, multi-scale, learning, lightweight