Akademik Veri Yönetim Sistemi
IX-INTERNATIONAL ANTALYA SCIENTIFIC RESEARCH AND INNOVATIVE STUDIES CONGRESS, Antalya, Türkiye, 20 - 23 Kasım 2025, ss.444-458, (Tam Metin Bildiri)
Breast cancer is the most common malignancy among women and the second most frequently diagnosed cancer worldwide. Early detection greatly improves treatment outcomes and survival, especially in localized stages. Conventional imaging modalities such as mammography, CT, MRI, and US are essential but limited by radiation exposure, high cost, and contrast agent requirements. These drawbacks have increased interest in cost-effective, non-invasive alternatives such as infrared thermal imaging (ITI), capable of detecting surface temperature variations linked to malignancy. This study proposes a deep learning framework that combines vessel enhancement–based preprocessing and transfer learning to optimize breast thermogram classification accuracy. In this study, images obtained from the DMR-IR dynamic dataset were used, divided into two subsets containing 1120 (DS-1) and 4500 (DS-2) thermograms. To enhance the visibility of vascular structures and regional temperature variations, a vessel enhancement–based preprocessing approach was implemented. The preprocessing process was quantitatively evaluated using NIQE, PIQE, and BRISQUE metrics to determine optimal parameters. The combined use of CLAHE, Gaussian Blur and Sobel filters yielded the best scores, producing clearer vessel boundaries and distinct thermal gradients. The vessel-enhanced images were then processed using transfer learning–based CNN models, including ResNet50, DenseNet121, VGG16, and VGG19, to extract the relevant features. The extracted features were subsequently fed into a custom-designed classifier to perform the final classification. Vessel enhancement–based preprocessing improved both image quality and classification accuracy. In DS-1, accuracy rose from 99.4% to 100% for ResNet50, from 96.11% to 98.89% for VGG16, and from 89.98% to 100% for VGG19. In DS-2, ResNet50 achieved the best performance, improving from 91.25% to 93.96%. These findings highlight the potential of vessel enhancement–based preprocessing as a practical strategy to improve the diagnostic accuracy and generalization of deep learning models in infrared breast thermography.