Scholarlink

<p>Skin cancer is the most common type of cancer globally, with over 1.5 million new cases diagnosed annually. Melanoma, a dangerous form of skin cancer, can be life-threatening if not diagnosed early, while nonmelanoma skin cancer is more common and usually treatable. Early detection is critical, but traditional diagnostic methods, such as visual inspection and biopsy, can be invasive, slow, and inconsistent across practitioners. To address these challenges, our research aims to develop an accurate melanoma detection system using deep learning techniques. Our study was designed to tackle key gaps in existing research, such as the lack of diverse and balanced datasets, inconsistent image preprocessing methods, and poor model performance on low-quality images. Using the HAM10k dataset, data augmentation was implemented to address dataset imbalances and experimented with various preprocessing techniques to standardize the input images. Our primary model is a MobileNet-v2 fine-tuned using transfer learning, optimized with Bayesian hyperparameter optimization and K-fold cross-validation. This model achieved an impressive accuracy of 94.19%. Additionally, a hybrid model was developed by combining Inception-v3 and MobileNet-v2, which achieved a slightly higher accuracy of 94.27%. These results demonstrate the potential of using pre- trained models and their combinations in improving skin cancer detection. In the future, our plan is to extend this work to multiclass skin cancer classification and integrate these models into a user-friendly web application, making early screening more accessible and efficient for a broader population.&#xa0;</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

Advancing melanoma detection with transfer learning and hybrid models

  • 期刊论文

  • 发表日期：

    2025年12月19日

查看全文
  • Ananya Gupta,
  • Kashish Bansal,
  • Arti,
  • Aditi Sabharwal,
  • S. R. N. Reddy,
  • Rishika Anand

摘要

Skin cancer is the most common type of cancer globally, with over 1.5 million new cases diagnosed annually. Melanoma, a dangerous form of skin cancer, can be life-threatening if not diagnosed early, while nonmelanoma skin cancer is more common and usually treatable. Early detection is critical, but traditional diagnostic methods, such as visual inspection and biopsy, can be invasive, slow, and inconsistent across practitioners. To address these challenges, our research aims to develop an accurate melanoma detection system using deep learning techniques. Our study was designed to tackle key gaps in existing research, such as the lack of diverse and balanced datasets, inconsistent image preprocessing methods, and poor model performance on low-quality images. Using the HAM10k dataset, data augmentation was implemented to address dataset imbalances and experimented with various preprocessing techniques to standardize the input images. Our primary model is a MobileNet-v2 fine-tuned using transfer learning, optimized with Bayesian hyperparameter optimization and K-fold cross-validation. This model achieved an impressive accuracy of 94.19%. Additionally, a hybrid model was developed by combining Inception-v3 and MobileNet-v2, which achieved a slightly higher accuracy of 94.27%. These results demonstrate the potential of using pre- trained models and their combinations in improving skin cancer detection. In the future, our plan is to extend this work to multiclass skin cancer classification and integrate these models into a user-friendly web application, making early screening more accessible and efficient for a broader population.