Scholarlink

<p>Analyzing chest X-rays is crucial for early lung disease detection, yet automatic detection remains challenging due to varying lesion sizes, complex anatomical structures, and subtle abnormalities. Existing models struggle with balancing accuracy and computational efficiency. This study proposes a novel deep learning architecture for detecting pulmonary abnormalities in chest X-rays from the VinBigData dataset. The model integrates an efficient REPNCSPELAN4 backbone with down-sampling layers for robust feature extraction. An improved SPPELAN module, combining spatial pyramid pooling and efficient layer aggregation network, enhances multi-scale feature learning and gradient flow. To mitigate information loss, a programmable gradient information technique is used, which preserves essential data across layers, optimizing detection performance. A generalized ELAN further improves parameter efficiency. Additionally, an anchor-free detection head eliminates predefined anchor boxes, enhancing accuracy while predicting lesions at three scales. Experimental results demonstrate superior performance over more recent state-of-the-art models, specifically YOLOv8, v9, v10 and transformer-based detectors such as RT-DETR, achieving a precision of 0.607, recall of 0.341, mAP@0.5 of 0.475, and mAP@0.5–0.95 of 0.298. The model also achieves an inference speed of 25.1&#xa0;ms, emphasizing computational efficiency and real-time applicability. This advancement supports automated, accurate, and scalable chest X-ray analysis, aiding radiologists in clinical decision-making.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Efficient Deep Learning Framework for Pulmonary Abnormality Detection in Chest X-Rays with Multi-Scale Feature Extraction and Anchor-Free Detection

  • 期刊论文

  • 发表日期：

    2025年9月20日

查看全文
  • Lakshmi Prasanna Kothala,
  • A. Nageswara Rao,
  • Joshua Reginald Pullagura,
  • Kiran Kumar Chanumolu,
  • Y. Chakrapani

摘要

Analyzing chest X-rays is crucial for early lung disease detection, yet automatic detection remains challenging due to varying lesion sizes, complex anatomical structures, and subtle abnormalities. Existing models struggle with balancing accuracy and computational efficiency. This study proposes a novel deep learning architecture for detecting pulmonary abnormalities in chest X-rays from the VinBigData dataset. The model integrates an efficient REPNCSPELAN4 backbone with down-sampling layers for robust feature extraction. An improved SPPELAN module, combining spatial pyramid pooling and efficient layer aggregation network, enhances multi-scale feature learning and gradient flow. To mitigate information loss, a programmable gradient information technique is used, which preserves essential data across layers, optimizing detection performance. A generalized ELAN further improves parameter efficiency. Additionally, an anchor-free detection head eliminates predefined anchor boxes, enhancing accuracy while predicting lesions at three scales. Experimental results demonstrate superior performance over more recent state-of-the-art models, specifically YOLOv8, v9, v10 and transformer-based detectors such as RT-DETR, achieving a precision of 0.607, recall of 0.341, mAP@0.5 of 0.475, and mAP@0.5–0.95 of 0.298. The model also achieves an inference speed of 25.1 ms, emphasizing computational efficiency and real-time applicability. This advancement supports automated, accurate, and scalable chest X-ray analysis, aiding radiologists in clinical decision-making.