Scholarlink

<p>Medical image segmentation is critical for accurate diagnosis and treatment planning, yet existing methods struggle with computational complexity and modality-specific challenges. We propose SC-TransED, a lightweight hybrid encoder–decoder framework that integrates a novel Channel-Spatial Transformer (CS-Tr) module to synergistically combine convolutional neural networks with transformer-based global context modeling. The CS-Tr module employs sequential channel attention, spatial attention, and transformer encoding to progressively refine features, enabling effective capture of both local texture details and long-range dependencies. To address ultrasound-specific challenges including speckle noise and boundary ambiguity, we introduce CS-TransED-Ultra, which incorporates a dedicated ultrasound-specific block with multiscale spatial filtering. Extensive experiments across five medical imaging modalities demonstrate state-of-the-art performance: 96.34% Dice score on BraTS 2018 (brain MRI), 96.49% on chest X-ray, 95.52% on PH2 (dermoscopy), 97.51% on Red Lesion (endoscopy), and 88.40% on BUSI (ultrasound with the ultrasound-specific variant). Despite achieving competitive performance, SC-TransED maintains exceptional computational efficiency with only 2&#xa0;M parameters and 5.73 GFLOPs. Ablation studies confirm the critical contribution of each architectural component, while systematic optimizer analysis using Q-Q plots, Dice scores, and pairwise t-tests identifies Adamax as the most effective optimizer (96.34% Dice), with statistical validation confirming significant performance differences across optimization strategies. By effectively balancing the complementary strengths of CNNs and transformers within a lightweight architecture, our framework addresses the critical challenges of computational complexity and scalability, offering a practical, deployable solution for AI-assisted diagnosis in resource-constrained clinical settings.</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

CS-TransED: An Optimized Hybrid CNN-Transformer with Channel-Spatial Attention for Computationally Efficient and High-Accuracy Multimodal Medical Image Segmentation

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Noura Bentaher,
  • Younes Kabbadj,
  • Samira Lafraxo,
  • Mohamed Ben Salah

摘要

Medical image segmentation is critical for accurate diagnosis and treatment planning, yet existing methods struggle with computational complexity and modality-specific challenges. We propose SC-TransED, a lightweight hybrid encoder–decoder framework that integrates a novel Channel-Spatial Transformer (CS-Tr) module to synergistically combine convolutional neural networks with transformer-based global context modeling. The CS-Tr module employs sequential channel attention, spatial attention, and transformer encoding to progressively refine features, enabling effective capture of both local texture details and long-range dependencies. To address ultrasound-specific challenges including speckle noise and boundary ambiguity, we introduce CS-TransED-Ultra, which incorporates a dedicated ultrasound-specific block with multiscale spatial filtering. Extensive experiments across five medical imaging modalities demonstrate state-of-the-art performance: 96.34% Dice score on BraTS 2018 (brain MRI), 96.49% on chest X-ray, 95.52% on PH2 (dermoscopy), 97.51% on Red Lesion (endoscopy), and 88.40% on BUSI (ultrasound with the ultrasound-specific variant). Despite achieving competitive performance, SC-TransED maintains exceptional computational efficiency with only 2 M parameters and 5.73 GFLOPs. Ablation studies confirm the critical contribution of each architectural component, while systematic optimizer analysis using Q-Q plots, Dice scores, and pairwise t-tests identifies Adamax as the most effective optimizer (96.34% Dice), with statistical validation confirming significant performance differences across optimization strategies. By effectively balancing the complementary strengths of CNNs and transformers within a lightweight architecture, our framework addresses the critical challenges of computational complexity and scalability, offering a practical, deployable solution for AI-assisted diagnosis in resource-constrained clinical settings.