MM-3D Unet: development of a lightweight breast cancer tumor segmentation network utilizing multi-task and depthwise separable convolution

Abstract

Background and objectives: This paper introduces a novel lightweight MM-3DUNet (Multi-task Mobile 3D UNet) network designed for efficient and accurate segmentation of breast cancer tumors masses from MRI images, which leverages depth-wise separable convolutions, channel expansion units, and auxiliary classification tasks to enhance feature representation and computational efficiency. Methods: We propose a 3D depth-wise separable convolution, and construct channel expansional convolution (CEC) unit and inverted residual block (IRB) to reduce the parameter count and computational load, making the network more suitable for use in resource-constrained environments. In addition, an auxiliary classification task (ACT) is introduced in the proposed architecture to provide additional supervisory signals for the main task of segmentation. The network architecture features a contracting path for downsampling and an expanding path for precise localization, enhanced by skip connections that integrate multi-level semantic information. Results: The network was evaluated using a dataset of Dynamic Contrast Enhanced MRI (DCE-MRI) breast cancer images, and the results show that compared to the classical 3DU-Net, MM-3DUNet could significantly reduce model parameters by 63.16% and computational demands by 80.90%, while increasing segmentation accuracy by 1.30% in IoU (Intersection over Union). Conclusions: MM-3DUNet offers a substantial reduction in computational requirements of breast cancer mass segmentation network. This network not only enhances diagnostic precision but also supports deployment in diverse clinical settings, potentially improving early detection and treatment outcomes for breast cancer patients.