Scholarlink

Background <p>Renal cancer is still a significant worldwide health issue where early detection and correct identification are crucial to enhancing patient survival. While 3D computed tomography (3D-CT) scans offer dense anatomical details for tumor evaluation, renal carcinomas are difficult to segment because of high intra-class variation, imaging artifacts, and noise contamination. Current deep learning solutions are generally not robust with adequate generalization, prone to overfitting with small sets, and overly computationally intensive, which hinders their use in the clinic.</p> Methods <p>To address these issues, the present work proposes a new deep learning architecture, Multi-Scale RenalNet-3D (MSRNet-3D), tailored for accurate renal tumor segmentation from 3D-CT scans. The method incorporates multi-scale and multi-level convolutional feature extraction for efficient capture of intricate tumor morphology, while the Seahorse Adaptive Search (SAS) optimization method automatically adjusts network hyperparameters, allowing for fast convergence and minimized training instability. </p> Result <p>Experimental assessments were performed utilizing a 3D-CT dataset consisting of 12,600 labeled renal cancer slices from 220 patients, split into training (70%), validation (15%), and testing (15%) sets. Dropout (rate = 0.3) and data augmentation regularization mechanisms were used to counteract overfitting threats. Empirical findings show that MSRNet-3D obtains a segmentation accuracy of 99.0%, precision of 98.9%, recall of 99.0%, and F1-score of 98.8%, outperforming current state-of-the-art models by 2.6–3.1% on average.</p> Conclusion <p>The model also displays better robustness to tumor shape variation and noise distortions, with computational efficiency enhanced by 18% over traditional 3D-CNN architectures. This study fills the important gap between high-resolution imaging and AI-based segmentation precision, delivering a scalable and consistent computer-aided renal cancer diagnosis and treatment planning tool.</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

Renal Cancer Detection and Segmentation Innovating MSRNet-3D and SAS Optimization for Precise Diagnosis with 3D-CT Imaging

  • 期刊论文

  • 发表日期：

    2026年2月18日

查看全文
  • Swapna Jegatheesan,
  • Roselin Kiruba Robinson

摘要

Background

Renal cancer is still a significant worldwide health issue where early detection and correct identification are crucial to enhancing patient survival. While 3D computed tomography (3D-CT) scans offer dense anatomical details for tumor evaluation, renal carcinomas are difficult to segment because of high intra-class variation, imaging artifacts, and noise contamination. Current deep learning solutions are generally not robust with adequate generalization, prone to overfitting with small sets, and overly computationally intensive, which hinders their use in the clinic.

Methods

To address these issues, the present work proposes a new deep learning architecture, Multi-Scale RenalNet-3D (MSRNet-3D), tailored for accurate renal tumor segmentation from 3D-CT scans. The method incorporates multi-scale and multi-level convolutional feature extraction for efficient capture of intricate tumor morphology, while the Seahorse Adaptive Search (SAS) optimization method automatically adjusts network hyperparameters, allowing for fast convergence and minimized training instability.

Result

Experimental assessments were performed utilizing a 3D-CT dataset consisting of 12,600 labeled renal cancer slices from 220 patients, split into training (70%), validation (15%), and testing (15%) sets. Dropout (rate = 0.3) and data augmentation regularization mechanisms were used to counteract overfitting threats. Empirical findings show that MSRNet-3D obtains a segmentation accuracy of 99.0%, precision of 98.9%, recall of 99.0%, and F1-score of 98.8%, outperforming current state-of-the-art models by 2.6–3.1% on average.

Conclusion

The model also displays better robustness to tumor shape variation and noise distortions, with computational efficiency enhanced by 18% over traditional 3D-CNN architectures. This study fills the important gap between high-resolution imaging and AI-based segmentation precision, delivering a scalable and consistent computer-aided renal cancer diagnosis and treatment planning tool.