Scholarlink

<p>Deep neural networks for medical imaging often lack explicit encoding of intrinsic geometric structure, limiting their sensitivity to tumour boundaries and local intensity variations. This work proposes a geomeTry-conditioned differential attention (TCDA) mechanism for multi-class brain tumour classification from MRI. Unlike conventional attention mechanisms that operate directly on raw feature maps, TCDA constructs attention inputs from structured geometric representations computed from per-channel L2-normalised backbone features via Riesz-transform-inspired multi-scale representations (σ = 1, 2, 4) in the Gaussian scale-space. In this representation, phase captures structural transitions, while amplitude encodes local intensity variations, and both are fused through learnable scale weights. A two-head, channel-wise differential attention formulation is introduced, enabling element-wise modulation of feature channels rather than scalar attention responses. Attention is further conditioned using orientation-aware temperature scaling derived from pooled orientation statistics of the geometric representation, along with a topology-motivated gating signal based on phase zero-crossing density. A residual bypass and stochastic depth (<i>p</i> = 0.1) are incorporated to stabilise optimisation and prevent over-reliance on geometric cues. Experiments on a public brain tumour MRI dataset, together with cross-dataset evaluation on an independent cohort, demonstrate that TCDA consistently improves classification performance across multiple CNN backbones while maintaining robust generalisation. The results indicate that geometry-conditioned differential attention is particularly effective for architectures that benefit from enhanced structural and intensity-aware feature modelling. The code can be reproduced and is available at the following link - <a href="https://github.com/Madhav1422/TCDA-Net/tree/main">https://github.com/Madhav1422/TCDA-Net/tree/main</a> .</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

TCDA-Net: geometry-conditioned differential attention using riesz-inspired multi-scale features for brain tumour classification

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • B. Madhavan,
  • B. Raghavan,
  • K. Parkavikathirvelu,
  • B. Rukshana,
  • R. Balasubramanian,
  • S. Venkatesh,
  • Rengarajan Amirtharajan

摘要

Deep neural networks for medical imaging often lack explicit encoding of intrinsic geometric structure, limiting their sensitivity to tumour boundaries and local intensity variations. This work proposes a geomeTry-conditioned differential attention (TCDA) mechanism for multi-class brain tumour classification from MRI. Unlike conventional attention mechanisms that operate directly on raw feature maps, TCDA constructs attention inputs from structured geometric representations computed from per-channel L2-normalised backbone features via Riesz-transform-inspired multi-scale representations (σ = 1, 2, 4) in the Gaussian scale-space. In this representation, phase captures structural transitions, while amplitude encodes local intensity variations, and both are fused through learnable scale weights. A two-head, channel-wise differential attention formulation is introduced, enabling element-wise modulation of feature channels rather than scalar attention responses. Attention is further conditioned using orientation-aware temperature scaling derived from pooled orientation statistics of the geometric representation, along with a topology-motivated gating signal based on phase zero-crossing density. A residual bypass and stochastic depth (p = 0.1) are incorporated to stabilise optimisation and prevent over-reliance on geometric cues. Experiments on a public brain tumour MRI dataset, together with cross-dataset evaluation on an independent cohort, demonstrate that TCDA consistently improves classification performance across multiple CNN backbones while maintaining robust generalisation. The results indicate that geometry-conditioned differential attention is particularly effective for architectures that benefit from enhanced structural and intensity-aware feature modelling. The code can be reproduced and is available at the following link - https://github.com/Madhav1422/TCDA-Net/tree/main .