Scholarlink

Infrared thermography (IRT) is a non-invasive imaging able to detect irregularities in the temperature distribution of the skin. Particularly, IRT could be effective to detect anomalies related to the presence of breast cancer through the employment of artificial intelligence (AI) approaches. However, in order to effectively transfer this technological achievement in clinical practice, some explainable AI (XAI) approaches should be integrated to foster the trust of clinicians towards these approaches and to provide explanations regarding the identification process performed by the machinery. The objective of this research is to develop an automated explainable deep learning system for identifying IRT images of breast cancer (i.e., benign and malignant). Thermal images served as input for various convolutional neural networks (CNNs), including both a customized version and those based on transfer learning (i.e., VGG-16 and ResNet50 models). The customized CNN achieved around 85% accuracy in distinguishing between breast images with benign and malignant tumors. Approaches utilizing transfer learning produced better results, achieving 94.5% accuracy with VGG-16 and 95.5% with ResNet50. All the models were associated with a heatmap produced through the grad-cam approach, and an overlay with the IRT images was performed in order to show the areas that mainly contributed to the classification. These results illustrate the effectiveness of merging IRT with AI techniques to accurately differentiate between benign and malignant breast cancer. Moreover, the XAI approach fosters the employment of IRT as a supplementary diagnostic resource in cancer care.

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

Explainable Transfer Learning Models for Thermal Imaging-Based Breast Cancer Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Samuele Salvati,
  • Daniele Sacripante,
  • Daniela Cardone,
  • David Perpetuini,
  • Arcangelo Merla

摘要

Infrared thermography (IRT) is a non-invasive imaging able to detect irregularities in the temperature distribution of the skin. Particularly, IRT could be effective to detect anomalies related to the presence of breast cancer through the employment of artificial intelligence (AI) approaches. However, in order to effectively transfer this technological achievement in clinical practice, some explainable AI (XAI) approaches should be integrated to foster the trust of clinicians towards these approaches and to provide explanations regarding the identification process performed by the machinery. The objective of this research is to develop an automated explainable deep learning system for identifying IRT images of breast cancer (i.e., benign and malignant). Thermal images served as input for various convolutional neural networks (CNNs), including both a customized version and those based on transfer learning (i.e., VGG-16 and ResNet50 models). The customized CNN achieved around 85% accuracy in distinguishing between breast images with benign and malignant tumors. Approaches utilizing transfer learning produced better results, achieving 94.5% accuracy with VGG-16 and 95.5% with ResNet50. All the models were associated with a heatmap produced through the grad-cam approach, and an overlay with the IRT images was performed in order to show the areas that mainly contributed to the classification. These results illustrate the effectiveness of merging IRT with AI techniques to accurately differentiate between benign and malignant breast cancer. Moreover, the XAI approach fosters the employment of IRT as a supplementary diagnostic resource in cancer care.