Scholarlink

<p>Prostate cancer remains a significant global health concern, characterized by challenges in early detection, precise tumor delineation, and diagnostic consistency. This study introduces TrionixNet, a novel tri-modular deep learning architecture tailored for prostate cancer segmentation using multiparametric MRI. TrionixNet incorporates several innovative components, including a dual-path T-block that combines standard and dilated convolutions for capturing both local texture and global context, and a dynamic channel attention mechanism applied throughout the encoder-decoder hierarchy for continuous feature recalibration.</p><p> A dedicated Boundary Refinement Block is introduced to enhance segmentation sharpness at tumor boundaries. Critically, the architecture integrates the novel N-Core Attention mechanism, a generalized multi-head projection approach that expands traditional self-attention by enabling inter-projection feature reasoning across multiple learned representations. This allows the model to interpret subtle and ambiguous tumor patterns from multiple semantic perspectives. Furthermore, multi-scale feature fusion aggregates information across different encoder depths, ensuring the preservation of both fine-grained and high-level features. Evaluated on a dataset of 680 MRI slices, TrionixNet significantly outperformed existing state-of-the-art models, achieving a Dice score of 0.9862, IoU of 0.9505, and Hausdorff distance of 0.049, alongside sensitivity (0.9761) and specificity (0.9999). Ablation studies validated the contribution of each architectural block. These results position TrionixNet as a powerful, interpretable, and clinically promising framework for automated prostate cancer diagnosis, with strong potential to accelerate progress in real-world precision oncology.</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 Comparative Evaluation of Deep Learning Architectures for Prostate Cancer Segmentation: Introducing TrionixNet with N-Core Multi-Attention Mechanism

  • 期刊论文

  • 发表日期：

    2025年10月12日

查看全文
  • Yogendra Narayan,
  • Davinder Paul Singh,
  • Tathagat Banerjee,
  • Pawandeep Kour,
  • Kantilal Rane,
  • Anand Deva Durai C,
  • K. Punnam Chandar,
  • Rajanish Kumar Kaushal,
  • Ram Murat Singh,
  • Yuvika Dhillon

摘要

Prostate cancer remains a significant global health concern, characterized by challenges in early detection, precise tumor delineation, and diagnostic consistency. This study introduces TrionixNet, a novel tri-modular deep learning architecture tailored for prostate cancer segmentation using multiparametric MRI. TrionixNet incorporates several innovative components, including a dual-path T-block that combines standard and dilated convolutions for capturing both local texture and global context, and a dynamic channel attention mechanism applied throughout the encoder-decoder hierarchy for continuous feature recalibration.

A dedicated Boundary Refinement Block is introduced to enhance segmentation sharpness at tumor boundaries. Critically, the architecture integrates the novel N-Core Attention mechanism, a generalized multi-head projection approach that expands traditional self-attention by enabling inter-projection feature reasoning across multiple learned representations. This allows the model to interpret subtle and ambiguous tumor patterns from multiple semantic perspectives. Furthermore, multi-scale feature fusion aggregates information across different encoder depths, ensuring the preservation of both fine-grained and high-level features. Evaluated on a dataset of 680 MRI slices, TrionixNet significantly outperformed existing state-of-the-art models, achieving a Dice score of 0.9862, IoU of 0.9505, and Hausdorff distance of 0.049, alongside sensitivity (0.9761) and specificity (0.9999). Ablation studies validated the contribution of each architectural block. These results position TrionixNet as a powerful, interpretable, and clinically promising framework for automated prostate cancer diagnosis, with strong potential to accelerate progress in real-world precision oncology.