A dual attention-guided 3D convolution network for automatic segmentation of prostate and tumor

Background: In middle-aged and older men, prostate cancer (PCa) is a common tumor disease with a mortality rate second only to lung cancer. The automatic and accurate segmentation of the prostate and tumor in magnetic resonance imaging (MRI) images can help doctors diagnose malignancies more efficiently. T2 weighted imaging (T2W) is now used in the majority of studies on prostate MRI image segmentation; however, diffusion-weighted imaging (DWI) is more valuable in the diagnosis of PCa. The morphological differences between the prostate and tumor regions are minimal, the tumor size is uncertain, the border between the tumor and surrounding tissue is hazy, and the categories separating normal regions from tumors are uneven. Consequently, it is challenging to segment prostate and tumor on DWI images.Methods: For the segmentation of prostate and tumor regions on DWI images, this study offers a dual attentionguided 3D convolutional neural network (3D DAG-Net). A visual attention method is built into the encoder step to obtain the features of various receptive fields and deliver more detailed contextual information. A multiscale attention technique is proposed at the decoder stage to fuse multiscale features to acquire finer global and local details. To resolve the class discrepancies between the prostate, tumor, and background regions in segmentation tasks, we propose a hybrid loss function for handling class imbalance.Results: We tested the algorithm on DWI images of PCa obtained from a nearby hospital, demonstrating the uniqueness and effectiveness of the method. Dice similarity coefficient (DSC) values for prostate and tumor DWI segmentation were 92.28% and 88.73%, respectively.Conclusion: We present a unique dual-attention mechanism 3D segmentation network architecture for quantitative assessment of prostate and tumor volumes on DWI. The automatic segmentation results produced by our technology were highly correlated and consistent with expert manual segmentation findings.