Scholarlink

Infrared thermography is a low-cost, non-invasive technique with potential for early breast cancer detection. However, clinical adoption is hindered by limited data availability and high interpatient variability. This work proposes a novel end-to-end framework that combines Generative Adversarial Networks (GANs) and Convolutional Neural Networks (CNNs) to generate synthetic thermal images, perform automated breast segmentation, and classify anomalous patterns. The Visual Lab Breast Thermography dataset was used with thermal normalization and automatic detection of the region of interest (ROI) based on morphological and thermal features. A CycleGAN architecture is trained to generate realistic images, evaluated using the Fréchet Inception Distance (FID) and Kernel Inception Distance (KID). Synthetic data were combined with real images to train the EfficientNet-B0 and ResNet50 classifiers (in permissive and strict configurations), achieving an accuracy of over 97% and an area under the ROC curve (AUC) of 0.977. The results demonstrate the feasibility of integrating GAN-based augmentation with CNN classification for robust automated breast screening. This method offers a promising diagnostic support system, especially suited for low-resource or mobile healthcare environments.

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 End-to-End GAN-CNN Framework for Early Breast Anomaly Detection Using Thermal Imaging

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hector Espinos-Morato,
  • Carlos Barroso,
  • Johanna Angulo,
  • Alex De María Espiritusanto,
  • Jenny Paula Aguilar Guiñez,
  • Cesáreo Tello Sánchez

摘要

Infrared thermography is a low-cost, non-invasive technique with potential for early breast cancer detection. However, clinical adoption is hindered by limited data availability and high interpatient variability. This work proposes a novel end-to-end framework that combines Generative Adversarial Networks (GANs) and Convolutional Neural Networks (CNNs) to generate synthetic thermal images, perform automated breast segmentation, and classify anomalous patterns. The Visual Lab Breast Thermography dataset was used with thermal normalization and automatic detection of the region of interest (ROI) based on morphological and thermal features. A CycleGAN architecture is trained to generate realistic images, evaluated using the Fréchet Inception Distance (FID) and Kernel Inception Distance (KID). Synthetic data were combined with real images to train the EfficientNet-B0 and ResNet50 classifiers (in permissive and strict configurations), achieving an accuracy of over 97% and an area under the ROC curve (AUC) of 0.977. The results demonstrate the feasibility of integrating GAN-based augmentation with CNN classification for robust automated breast screening. This method offers a promising diagnostic support system, especially suited for low-resource or mobile healthcare environments.