Scholarlink

<p>Skin cancer is considered a highly dangerous disease among other cancers. The rapid propagation ability of this disease makes skin cancer more fatal and dangerous. Therefore, the timely recognition of skin cancer is very important. Nevertheless, the diagnosis of skin cancer via dermoscopic imaging encounters persistent complexities, including image noise, changes in lesion appearance, high misclassification rates, and inaccurate segmentation. These problems often lead to incorrect or delayed diagnoses, limiting the efficacy of conventional deep-learning approaches. To resolve these problems, this research recommends a robust automated deep learning approach that incorporates improved heuristic optimization for improving the segmentation and classification performance. Firstly, the significant raw images are aggregated from the standard datasets. These images are further subjected to the pre-processing stage. This pre-processing phase prevents unwanted noise and outliers from raw images. It improves the efficiency of the diagnosis process and minimizes the computation time. After obtaining the pre-processed images, the next step involves image segmentation. The segmentation process&#xa0;is done by adaptive TransUNet, whereas the parameters of TransUNet are optimized by the developed Probability Ratio-based Reptile Hybrid Leader Optimizer (PR-RHLO)&#xa0;algorithm. In this segmentation process, the suggested adaptive TransUNet process efficiently segments the abnormal lesions, thus helping in the classification process. Here, the PR-RHLO increases the performance rates of the segmentation process by tuning the TransUNet technique parameters. Consequently, the classification is performed by a Parameter Tuned Residual Gated Attention Network (PT-RGAN), in which the hyper-parameters of the Residual network (ResNet) and Gated attention mechanism are optimally tuned using the recommended PR-RHLO. The suggested PR-RGAN is an effective and powerful model for categorizing skin cancer and minimizes the misclassification rates. Thus, this model minimizes the mortality rates and enhances the early treatment process. The final outcomes are achieved by taking the average among the two networks. At last, the performance is measured and validated through several parameter metrics, which are then distinguished from other traditional models. Therefore, the higher accuracy value of the suggested model is achieved as 96.9% for Dataset 1, 97% for Dataset 2, and 97.1% for Dataset 3, respectively. Therefore, the results indicate that the designed approach provides a powerful diagnostic tool that improves timely identification, minimizes mortality risk, and supports dermatologists in clinical decision-making with higher efficiency and reliability.</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

An Efficient Framework of Skin Lesion Segmentation and Classification Using Hybrid Heuristic Approach-Aided TransUNet and Residual Gated Attention Network

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • V. Chakkarapani,
  • S. Poornapushpakala

摘要

Skin cancer is considered a highly dangerous disease among other cancers. The rapid propagation ability of this disease makes skin cancer more fatal and dangerous. Therefore, the timely recognition of skin cancer is very important. Nevertheless, the diagnosis of skin cancer via dermoscopic imaging encounters persistent complexities, including image noise, changes in lesion appearance, high misclassification rates, and inaccurate segmentation. These problems often lead to incorrect or delayed diagnoses, limiting the efficacy of conventional deep-learning approaches. To resolve these problems, this research recommends a robust automated deep learning approach that incorporates improved heuristic optimization for improving the segmentation and classification performance. Firstly, the significant raw images are aggregated from the standard datasets. These images are further subjected to the pre-processing stage. This pre-processing phase prevents unwanted noise and outliers from raw images. It improves the efficiency of the diagnosis process and minimizes the computation time. After obtaining the pre-processed images, the next step involves image segmentation. The segmentation process is done by adaptive TransUNet, whereas the parameters of TransUNet are optimized by the developed Probability Ratio-based Reptile Hybrid Leader Optimizer (PR-RHLO) algorithm. In this segmentation process, the suggested adaptive TransUNet process efficiently segments the abnormal lesions, thus helping in the classification process. Here, the PR-RHLO increases the performance rates of the segmentation process by tuning the TransUNet technique parameters. Consequently, the classification is performed by a Parameter Tuned Residual Gated Attention Network (PT-RGAN), in which the hyper-parameters of the Residual network (ResNet) and Gated attention mechanism are optimally tuned using the recommended PR-RHLO. The suggested PR-RGAN is an effective and powerful model for categorizing skin cancer and minimizes the misclassification rates. Thus, this model minimizes the mortality rates and enhances the early treatment process. The final outcomes are achieved by taking the average among the two networks. At last, the performance is measured and validated through several parameter metrics, which are then distinguished from other traditional models. Therefore, the higher accuracy value of the suggested model is achieved as 96.9% for Dataset 1, 97% for Dataset 2, and 97.1% for Dataset 3, respectively. Therefore, the results indicate that the designed approach provides a powerful diagnostic tool that improves timely identification, minimizes mortality risk, and supports dermatologists in clinical decision-making with higher efficiency and reliability.