Scholarlink

<p>The reliable identification of pulmonary nodules at an early stage remains a significant challenge because of variations in nodule appearance, small lesion size, and frequent false positives in computed tomography (CT) analysis. This study presents a unified computer-aided diagnosis (CAD) framework that integrates image enhancement, segmentation, classification, and parameter optimization into a single automated pipeline. For image quality enhancement, the contrast and illumination were normalized using Multi-Scale Retinex with Color Restoration and Contrast-Limited Adaptive Histogram Equalization. Segmentation was performed using a Cross-Attention Dense U-Net++ (CADU-Net++), which extends the standard U-Net + + by incorporating cross-attention connections between the encoder and decoder layers to retain fine boundary and contextual features. Nodule classification is performed using a transformer-based 3D Dual-Path Network (3D-TDPN) that combines convolutional and transformer representations to capture both local and global spatial dependencies. The model hyperparameters were optimized using a Hybrid Quantum-Inspired Firefly–Whale Optimization Algorithm (HQFWOA), which merges the exploratory strength of the Firefly Algorithm and the exploitative mechanism of the Whale Optimization Algorithm through a quantum-probabilistic update process to achieve faster convergence and improved generalization. The proposed framework was evaluated on two benchmark datasets, LUNA16 and LIDC-IDRI, using a patient-wise 70:30 training-to-testing split and five-fold cross-validation to ensure statistical reliability and reproducibility. On the LUNA16 dataset, the model achieved an average Intersection over Union (IoU) of 96.80 ± 0.14 and a classification accuracy of 98.50 ± 0.20. On the LIDC-IDRI dataset, the model achieved an IoU of 97.25 ± 0.12 and a classification accuracy of 98.90 ± 0.18.These results consistently surpassed recent baselinesby 2–4% in segmentation performance and 1–2% in classification metrics across both the datasets. The findings demonstrate that the integration of cross-attention dense connectivity and hybrid metaheuristic optimization provides a novel, statistically validated, and generalizable approach for highly accurate early stage lung nodule detection.</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

A hybrid deep learning framework for automated lung nodule detection and classification using CADU-Net + + and transformer-based 3D dual-path network

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • M. Dhayalini,
  • Revathi alias Ponmozhi Balasundaram

摘要

The reliable identification of pulmonary nodules at an early stage remains a significant challenge because of variations in nodule appearance, small lesion size, and frequent false positives in computed tomography (CT) analysis. This study presents a unified computer-aided diagnosis (CAD) framework that integrates image enhancement, segmentation, classification, and parameter optimization into a single automated pipeline. For image quality enhancement, the contrast and illumination were normalized using Multi-Scale Retinex with Color Restoration and Contrast-Limited Adaptive Histogram Equalization. Segmentation was performed using a Cross-Attention Dense U-Net++ (CADU-Net++), which extends the standard U-Net + + by incorporating cross-attention connections between the encoder and decoder layers to retain fine boundary and contextual features. Nodule classification is performed using a transformer-based 3D Dual-Path Network (3D-TDPN) that combines convolutional and transformer representations to capture both local and global spatial dependencies. The model hyperparameters were optimized using a Hybrid Quantum-Inspired Firefly–Whale Optimization Algorithm (HQFWOA), which merges the exploratory strength of the Firefly Algorithm and the exploitative mechanism of the Whale Optimization Algorithm through a quantum-probabilistic update process to achieve faster convergence and improved generalization. The proposed framework was evaluated on two benchmark datasets, LUNA16 and LIDC-IDRI, using a patient-wise 70:30 training-to-testing split and five-fold cross-validation to ensure statistical reliability and reproducibility. On the LUNA16 dataset, the model achieved an average Intersection over Union (IoU) of 96.80 ± 0.14 and a classification accuracy of 98.50 ± 0.20. On the LIDC-IDRI dataset, the model achieved an IoU of 97.25 ± 0.12 and a classification accuracy of 98.90 ± 0.18.These results consistently surpassed recent baselinesby 2–4% in segmentation performance and 1–2% in classification metrics across both the datasets. The findings demonstrate that the integration of cross-attention dense connectivity and hybrid metaheuristic optimization provides a novel, statistically validated, and generalizable approach for highly accurate early stage lung nodule detection.