Browsing by Author "Cihan, M."

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • Thumbnail Image
    Conference Object
    Multi-View Thermal Breast Imaging for Malignancy Detection: Performance Benchmarking of CNN, Transformer, and Involution Architectures
    (Springer Science and Business Media Deutschland GmbH, 2026) Cihan, M.; Ceylan, M.
    Breast cancer screening demands accurate, non-invasive, low-cost tools. Infrared thermography is radiation-free and portable, but its utility hinges on robust computer-aided diagnosis (CAD). We benchmark three deep-learning families for static multi-view breast thermography—CNNs, Transformers, and an involution-based model (HarmonyNet-Lite). Experiments use the Breast Thermography dataset (119 patients; 476 manually segmented ROIs from anterior/oblique views). A compact pipeline performs ROI segmentation, RGB conversion, normalization, resizing, and moderate data augmentation; class imbalance is handled with minority oversampling and class-weighted loss. Evaluation follows patient-stratified five-fold cross-validation. HarmonyNet-Lite yields the best results: accuracy 87.47 ± 2.99%, recall 93.33 ± 2.13%, F1 68.43 ± 8.75%, and precision 54.23 ± 8.94%, indicating high sensitivity with an acceptable trade-off in precision for screening. Among CNNs, ResNet50 is strongest (85.59 ± 3.37%; F1 63.16 ± 3.87%), followed by InceptionV3 (83.38 ± 1.41%; F1 59.99 ± 6.72%), while DenseNet121 lags (79.25 ± 2.98%; F1 52.38 ± 5.62%). Transformer performance is mixed: ViT-Tiny is competitive (84.59 ± 4.23%; F1 59.46 ± 4.68%), whereas Swin-Tiny trails (81.30 ± 2.32%; F1 57.14 ± 4.44%) due to lower precision. Despite using only 0.14 M parameters, HarmonyNet-Lite outperforms heavier CNNs (ResNet50: 23.59 M; InceptionV3: 21.81 M) and lighter Transformers (ViT-Tiny: 2.84 M; Swin-Tiny: 11.78 M), demonstrating that content-adaptive, spatially aware involution operators efficiently capture fine thermal gradients. These findings position HarmonyNet-Lite as a strong, deployable CAD candidate. Future work will pursue multi-center validation, automated segmentation, multi-class labeling, hybrid involution–attention/multimodal models, and controlled GAN-based augmentation to mitigate data scarcity. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.