Scholarlink

<p>The global rise in colorectal cancer mortality highlights the urgent need for early and accurate detection to improve treatment outcomes and patient survival. Multilevel thresholding is a widely used technique for extracting Regions of Interest (ROIs) in colorectal cancer images (polyps), as it enables precise segmentation by distinguishing tissue types based on intensity variations. However, its performance depends critically on selecting optimal thresholds, which remains challenging due to the heterogeneous appearance and complex structures of polyps. To overcome this, we propose GQAISA_Nstep (Gradient Q-learning Adolescent Identity Search Algorithm with N-step), an enhanced metaheuristic that integrates the Gradient Search Rule (GSR) with a multi-agent asynchronous N-step Q-learning controller. This hybrid design adaptively balances exploration and exploitation, enhances accuracy and stability, and ensures long-term learning through uniform-mixture updates. Furthermore, three tailored mutation strategies are incorporated to enrich population diversity, mitigate premature convergence, and improve robustness against local optima. The effectiveness of GQAISA_Nstep is rigorously validated on two datasets. On the BSDS500 benchmark dataset, the algorithm demonstrated a superior ability to escape local minima. On the PolypDB dataset, used for colorectal cancer polyp segmentation, GQAISA_Nstep consistently outperformed baseline and advanced metaheuristics, achieving an average QILV of 0.98, PSNR of 27.5 dB, and RMSE of 10.7. Strong results were also observed across 12 evaluation metrics, including FSIM, SSIM, AD, MD, NAE, UIQI, HPSI, NCC, RMSE, PSNR, QILV, run time and fitness, confirming its robustness and accuracy. These findings highlight GQAISA_Nstep as a promising tool for computer-aided colorectal cancer diagnosis.</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

Colorectal polyp segmentation using an adolescent identity search algorithm with gradient Q-learning and asynchronous N-step updates

  • 期刊论文

  • 发表日期：

    2026年5月23日

查看全文
  • Amir Hamza,
  • Yassine Himeur,
  • Badis Lekouaghet,
  • Morad Grimes,
  • Abdelkarim Boukabou,
  • Adel Oulefki

摘要

The global rise in colorectal cancer mortality highlights the urgent need for early and accurate detection to improve treatment outcomes and patient survival. Multilevel thresholding is a widely used technique for extracting Regions of Interest (ROIs) in colorectal cancer images (polyps), as it enables precise segmentation by distinguishing tissue types based on intensity variations. However, its performance depends critically on selecting optimal thresholds, which remains challenging due to the heterogeneous appearance and complex structures of polyps. To overcome this, we propose GQAISA_Nstep (Gradient Q-learning Adolescent Identity Search Algorithm with N-step), an enhanced metaheuristic that integrates the Gradient Search Rule (GSR) with a multi-agent asynchronous N-step Q-learning controller. This hybrid design adaptively balances exploration and exploitation, enhances accuracy and stability, and ensures long-term learning through uniform-mixture updates. Furthermore, three tailored mutation strategies are incorporated to enrich population diversity, mitigate premature convergence, and improve robustness against local optima. The effectiveness of GQAISA_Nstep is rigorously validated on two datasets. On the BSDS500 benchmark dataset, the algorithm demonstrated a superior ability to escape local minima. On the PolypDB dataset, used for colorectal cancer polyp segmentation, GQAISA_Nstep consistently outperformed baseline and advanced metaheuristics, achieving an average QILV of 0.98, PSNR of 27.5 dB, and RMSE of 10.7. Strong results were also observed across 12 evaluation metrics, including FSIM, SSIM, AD, MD, NAE, UIQI, HPSI, NCC, RMSE, PSNR, QILV, run time and fitness, confirming its robustness and accuracy. These findings highlight GQAISA_Nstep as a promising tool for computer-aided colorectal cancer diagnosis.