A Hybrid Multi-Model Deep Learning Framework for Breast Cancer Detection Using Thermogram Imagery

Abstract

Breast cancer is one of the most common diseases among women worldwide, which causes high mortality in case of late detection. Thermography is a non-invasive imaging technology that can be used to help with the early diagnosis of breast cancer. With the development of AI, thermogram-based breast cancer screening using deep learning techniques has gained significant value. However, the detection accuracy and robustness of current deep learning algorithms are still challenged. To resolve this, we designed a hybrid multi-model deep learning framework which combines ROI segmentation by ROISegNet, feature extraction with enhanced edge using two edge detectors (Prewits and Roberts), and finally feature extraction and classification by a deep network architecture, InceptionResNetV2. The proposed framework was tested on the DMR-IR dataset, which is publicly available and includes thermograms of 44 subjects (29 breast cancer patients and 15 healthy). The dataset was separated into training (580 affected and 300 healthy images) and validation (160 affected and 80 healthy images) sets. The proposed model also achieved better results than the state-of-the-art 98.78% accuracy, 97.97% precision, 96.52% recall, and 97.24% F1-score, compared with other base networks, such as VGG19, ResNet50, DenseNet121, and InceptionV3. This performance also indicates that the incorporation of edge-enhanced feature maps and ROI segmentation in a hybrid deep learning framework is a practical design. The method presented here represents a potential direction for reliable non-invasive early detection of breast cancer based on thermogram images.