3DSC2-UNet: Advancing Breast Cancer Segmentation Techniques Using the 3D Skipped Convolution Capsule-UNet

Breast cancer (BC) persists as a substantial global health hurdle, demanding ongoing progress in the realms of detection, diagnosis, and therapeutic approaches. This thorough examination delves into cutting-edge technologies and methodologies that have surfaced in the quest for enhanced results in BC management. Yet, investigations employing solitary imaging modalities have showcased restricted clinical effectiveness owing to their relatively diminished diagnostic precision and dependability when compared to multimodal systems. As a result, our aim is to create a 3D Skipped Convolution Capsule UNet designed for the segmentation of breast tumors, employing both mammogram and ultrasound imaging modalities. For the integration of these modalities, we utilize the Dynamic Fusion Net (DF-Net), incorporating two modules named the Molecular-Level Aligner Module and Biphasic Sequential Integration Module (BSInU). The process of segmentation takes place within our innovative framework, 3DSC2-UNet, encompassing four primary elements: (a) Graphical feature extraction, (b) Convolutional Capsule (ConvCaps) encoder module, (c) Synaptic skip connection, and (d) Convolutional decoder (Conv Decoder). The overall accuracy attained is 99.97%, accompanied by a Dice Similarity Coefficient (DSC) of 0.89 and an Average Surface Distance (ASD) of 2.5mm across the mammogram dataset. Moreover, the precision stands at 99.65%, accompanied by a Dice Similarity Coefficient (DSC) of 0.88 and an Average Surface Distance (ASD) of 2.4mm across the ultrasound dataset. The suggested 3DSC2-UNet, harnessing both mammogram and ultrasound, showcases exceptional performance in accuracy, ASD, and DSC metrics. This model holds the capability to diminish unwarranted biopsy procedures, thereby advancing its prospects for clinical implementation.