Scholarlink

Background <p>Medivision is a modern diagnostic system that has been developed to improve colorectal cancer detection.&#xa0;</p> Methodology <p>By applying deep learning acquisitions, including convolutional neural networks, Gray-Level Co-occurrence Matrix feature extraction, and visualization with Grad-CAM. After developing the Medivision system, several colonoscopy image datasets were used for evaluating a number of state-of-the-art deep CNN architectures, such as ResNet50, VGG16, VGG19, and DenseNet201—namely, CVC Clinic DB, Kvasir2, and Hyper Kvasir, however, through exhaustive model comparisons with integrated CNNs DEV-22 and RV-22. By the incorporation of GLCM feature extraction, textual analysis has become enhanced in the capturing of some critical features within an image that enhances model sensitivity in the detection of subtle variation within colorectal polyps. Grad-CAM visualizations enhance interpretability by enabling clinicians to attribute diagnostically important regions and insight into the model’s process of decision-making, incorporated with cloud storage.&#xa0;</p> Results <p>This work confirmed the stability and reliability of VGG16 in different conditions of image shooting. Among these, the best performances for all datasets were obtained by the model VGG16, confirming its consistency and robustness in varying conditions of image acquisition. On the CVC Clinic DB, VGG16 assured a training accuracy of 99.22% and a testing accuracy of 96.12%. The model, when trained on the Kvasir2 dataset, achieved an accuracy of 96.56% during training and 94.25% when testing, while on the Hyper Kvasir dataset, the model supported training accuracy of 95.19% and a testing accuracy of 98.87%. Apart from that, VGG16 also showed a good localization capability for the average IoU of 0.78 on CVC Clinic DB, 0.79 on Kvasir2, and 0.77 on Hyper Kvasir, indicating its precision in identifying polyp regions. It also tested the performance of integrated CNN models DEV-22 and RV-22 in complex multi-dataset scenarios. DEV-22 gave the best-performing integrated model against test accuracies of 97.86% against CVC Clinic DB, 89.37% against Kvasir2, and 76.08% against Hyper Kvasir, while RV-22 resulted in relatively poor performance across these datasets.This indicates that DEV-22 might be much better suited for colorectal cancer detection tasks whenever multiple datasets are concerned. The high testing accuracy and precise localization capabilities of VGG16 and DEV-22 show their robustness within the Medivision system for delivering appropriate clinically relevant returns on colorectal cancer screening across all three datasets.&#xa0;</p> Clinical relevance <p>Medivision enables real-time analysis in an accessible and efficient way for healthcare providers. This paper presents the clinical relevance of this system in enhancing diagnostic accuracy and interpretation for colorectal cancer screening workflows. </p> Conclusion <p>It represents a state-of-the-art integration of high-performance deep CNN models, GLCM, and Grad-CAM for accurate, interpretable, and actionable outcomes in the diagnosis of colorectal cancer, by representing one great bound in AI-mediated medical 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

Medivision: Empowering Colorectal Cancer Diagnosis and Tumor Localization Through Supervised Learning Classifications and Grad-CAM Visualization of Medical Colonoscopy Images

  • 期刊论文

  • 发表日期：

    2025年3月21日

查看全文
  • Akella S. Narasimha Raju,
  • K Venkatesh,
  • Ranjith Kumar Gatla,
  • Shaik Jakeer Hussain,
  • Subba Rao Polamuri

摘要

Background

Medivision is a modern diagnostic system that has been developed to improve colorectal cancer detection. 

Methodology

By applying deep learning acquisitions, including convolutional neural networks, Gray-Level Co-occurrence Matrix feature extraction, and visualization with Grad-CAM. After developing the Medivision system, several colonoscopy image datasets were used for evaluating a number of state-of-the-art deep CNN architectures, such as ResNet50, VGG16, VGG19, and DenseNet201—namely, CVC Clinic DB, Kvasir2, and Hyper Kvasir, however, through exhaustive model comparisons with integrated CNNs DEV-22 and RV-22. By the incorporation of GLCM feature extraction, textual analysis has become enhanced in the capturing of some critical features within an image that enhances model sensitivity in the detection of subtle variation within colorectal polyps. Grad-CAM visualizations enhance interpretability by enabling clinicians to attribute diagnostically important regions and insight into the model’s process of decision-making, incorporated with cloud storage. 

Results

This work confirmed the stability and reliability of VGG16 in different conditions of image shooting. Among these, the best performances for all datasets were obtained by the model VGG16, confirming its consistency and robustness in varying conditions of image acquisition. On the CVC Clinic DB, VGG16 assured a training accuracy of 99.22% and a testing accuracy of 96.12%. The model, when trained on the Kvasir2 dataset, achieved an accuracy of 96.56% during training and 94.25% when testing, while on the Hyper Kvasir dataset, the model supported training accuracy of 95.19% and a testing accuracy of 98.87%. Apart from that, VGG16 also showed a good localization capability for the average IoU of 0.78 on CVC Clinic DB, 0.79 on Kvasir2, and 0.77 on Hyper Kvasir, indicating its precision in identifying polyp regions. It also tested the performance of integrated CNN models DEV-22 and RV-22 in complex multi-dataset scenarios. DEV-22 gave the best-performing integrated model against test accuracies of 97.86% against CVC Clinic DB, 89.37% against Kvasir2, and 76.08% against Hyper Kvasir, while RV-22 resulted in relatively poor performance across these datasets.This indicates that DEV-22 might be much better suited for colorectal cancer detection tasks whenever multiple datasets are concerned. The high testing accuracy and precise localization capabilities of VGG16 and DEV-22 show their robustness within the Medivision system for delivering appropriate clinically relevant returns on colorectal cancer screening across all three datasets. 

Clinical relevance

Medivision enables real-time analysis in an accessible and efficient way for healthcare providers. This paper presents the clinical relevance of this system in enhancing diagnostic accuracy and interpretation for colorectal cancer screening workflows.

Conclusion

It represents a state-of-the-art integration of high-performance deep CNN models, GLCM, and Grad-CAM for accurate, interpretable, and actionable outcomes in the diagnosis of colorectal cancer, by representing one great bound in AI-mediated medical diagnosis.