A multi-scale attention-based parallel feature calibration framework for segmentation of multi-modality breast cancer images

Breast Cancer (BC) segmentation remains a challenging task because of the different possible appearances of the tumors or infected regions in different imaging modalities such as Ultrasound, Mammography, and Magnetic Resonance Imaging (MRI). The intensity, shape, and size variations, as well as other regions of the lesions in a single image, make the segmentation process challenging. In the presented study, a multi-Scale parallel feature calibration-based model is introduced with two specially designed modules, namely the Multi-scale Channel Attention Module (MCAM), which aims at the informative channels for the improvement of boundary detection, and Progressive Feature Calibration (PFC), which combines the multi-scale contextual description with the preservation of fine-grained details. The proposed encoder-decoder network is incorporated with MCAM and PFC in four hierarchical stages. To validate the effectiveness of the presented network, extensive experiments were conducted across three modalities, including Ultrasound, MRI, and Mammography. The proposed network achieved the Dice, Intersection over Union (IoU), and precision of 90.79%, 82.47%, and 95.78% on the Ultrasound dataset, 89.61%, 82.81%, and 95.47% on the MRI dataset, and 71.40%, 56.52%, and 86.33% on the mammography dataset, respectively. The obtained results demonstrate that the inclusion of MCAM and PFC significantly enhances the segmentation measures. Overall, the presented study contributes to identifying complex tumor segmentation regions across different imaging modalities.