BMA-Net: A 3D bidirectional multi-scale feature aggregation network for prostate region segmentation

Accurate segmentation of the prostate region in magnetic resonance imaging (MRI) is crucial for prostate-related diagnoses. Recent studies have incorporated Transformers into prostate region segmentation to better capture long-range global feature representations. However, due to the computational complexity of Transformers, these studies have been limited to processing single slices. Incorporating multiple slices can facilitate more precise segmentation, but existing methods fail to effectively utilize both intra-slice and inter-slice multi-scale information.

To address these challenges, we propose a 3D bidirectional multi-scale feature aggregation network, called BMA-Net. This network employs a forward frequency-based global feature filtering branch to learn and filter highly correlated information both intra-slice and inter-slice. It also includes a reverse spatial attention branch, guided by Gaussian distance, to model spatial information within slices. Additionally, a convolutional neural network (CNN) branch is incorporated to supplement local feature information. To mitigate feature discrepancies among different branches, the network uses a multi-scale feature fusion module for feature interaction.

Experiments on both public and in-house datasets were conducted. The results on the public dataset showed a Dice coefficient of 88.35 % in the central gland and 76.86 % in the peripheral zone. On the in-house dataset, the Dice coefficients were 85.85 % for the central gland and 74.50 % for the peripheral zone.

BMA-Net leverages multi-scale information both intra-slice and inter-slice to achieve more accurate segmentation of prostate regions. The experimental results demonstrate that our approach achieves superior segmentation performance compared to the current state-of-the-art methods.