Breast Tumor-Like-Masses Segmentation From Scattering Images Obtained With an Ultrahigh-Sensitivity Talbot-Lau Interferometer Using Convolutional Neural Networks

Segmentation of breast tumors using imaging techniques remains a critical challenge in medical diagnostics, with limitations in contrast and resolution affecting early detection. Conventional mammography methods still have some limitations in differentiating between healthy and tumorous tissue, particularly in cases with low-density differences. This study investigates the potential of combining ultrahigh-sensitivity Talbot-Lau interferometry with Convolutional Neural Networks (CNNs) to enhance breast tumor segmentation from scattering images. The research aims to improve tumor segmentation accuracy and efficiency by leveraging phase contrast imaging. The experimental setup utilized an ultrahigh-sensitivity Talbot-Lau interferometer operated with a conventional X-ray tube to generate scattering images, which were processed using a Fourier Transform-based algorithm. Five CNN architectures - U-Net, ResNet50, DeepLabV3, PSPNet, and SegNet -were trained and tested on an augmented dataset of 320 images. Performance was evaluated based on accuracy, precision, specificity, recall, and F1-score. DeepLabV3 achieved the highest accuracy (87.07%) and F1-score (90.92%), followed by PSPNet with 85.94% accuracy. However, significant fluctuations were observed in validation accuracy, indicating sensitivity to dataset variability. U-Net demonstrated the most stable performance with an accuracy of 86.34% and an F1-score of 90.2%, making it the most reliable model for tumor segmentation in scattering images. The combination of Talbot-Lau interferometry with CNNs presents a promising approach for breast tumor segmentation. U-Net emerged as the most stable model, suggesting its potential application in medical diagnostics. Future work will focus on optimizing CNN architecture and expanding the dataset to improve the segmentation of small tumor-like masses.