Scholarlink

<p>Skin cancer, particularly melanoma, is an aggressive form of cancer. The problem with human diagnosis is that it is too subjective and inefficient for detecting early signs. Artificial intelligence (AI) is a solution that can clearly be a very efficient, fast, objective, and accurate pathway to improve early interventions and patient care. A new hybrid model, MobileNetV2-ResNet101-ViTE, was developed, which effectively combines Convolutional Neural Network and the Vision Transformer Encoder (ViTE) using a Spatial Detail Enhancement Block. The proposed model differs from existing models by retaining accurate local histological representations and utilizing the global context necessary for accurate classification of skin lesions. Model performance was assessed on the ISIC2019 dataset, yielding an average Area Under the ROC Curve (AUC) of 96.94% and a classification accuracy of 98%. The model’s high scores in sensitivity to the malignant classes have astounding potential in improving false negatives with scores of 98.5% for Melanoma, and 91.5% for squamous cell carcinoma, giving some confidence in its potential to improve false negatives, and has consistently reported very high specificity (average: 99.55%) and precision (average: 94.44%). The proposed MobileNetV2-ResNet101-ViTE model is ground-breaking for dermatologists using an AI-driven approach for diagnosis and should provide a better outcome for early melanoma detection.</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

ViTE-MobileNetV2-ResNet101: Fusion Vision Transformer Encoder and CNNs Based on Spatial Detail Enhancement for Early Diagnosis Skin Cancer

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月10日

查看全文
  • Aisha M. Mashraqi,
  • Ebrahim Mohammed Senan,
  • Yousef Asiri,
  • Ibrahim Abunadi,
  • Hanan T. Halawani,
  • Eman A. Alshari

摘要

Skin cancer, particularly melanoma, is an aggressive form of cancer. The problem with human diagnosis is that it is too subjective and inefficient for detecting early signs. Artificial intelligence (AI) is a solution that can clearly be a very efficient, fast, objective, and accurate pathway to improve early interventions and patient care. A new hybrid model, MobileNetV2-ResNet101-ViTE, was developed, which effectively combines Convolutional Neural Network and the Vision Transformer Encoder (ViTE) using a Spatial Detail Enhancement Block. The proposed model differs from existing models by retaining accurate local histological representations and utilizing the global context necessary for accurate classification of skin lesions. Model performance was assessed on the ISIC2019 dataset, yielding an average Area Under the ROC Curve (AUC) of 96.94% and a classification accuracy of 98%. The model’s high scores in sensitivity to the malignant classes have astounding potential in improving false negatives with scores of 98.5% for Melanoma, and 91.5% for squamous cell carcinoma, giving some confidence in its potential to improve false negatives, and has consistently reported very high specificity (average: 99.55%) and precision (average: 94.44%). The proposed MobileNetV2-ResNet101-ViTE model is ground-breaking for dermatologists using an AI-driven approach for diagnosis and should provide a better outcome for early melanoma detection.