Automatic Segmentation of Rectal Tumor on Magnetic Resonance Images via a Deep Discriminative Model Consisting of Asymmetric U-Net and Convolutional Conditional Random Field based Post-processing
Xing Li, Hao Chen, Sisi Ge · 2023
The process of manual delineating is frequently time-consuming and can result in low consistency. Our goal was to design a deep discriminative model (DDM) to mitigate these issues of magnetic resonance imaging (MRI) for rectal tumor lesion boundary delineation, and we proposed an asymmetric U-Net rectal tumor MR images segmentation algorithm based on multi-scale dilated convolutional inputs and convolutional conditional random fields (ConvCRFs). A multiscale convolutional input module was designed as a preprocessing step to enrich the extraction and input of global contextual semantic information; then, an asymmetric U-Net network combined with ConvCRFs was used to discriminatively fine-tune the segmentation results so as to enhance segmentation accuracy of tumors. To verify the feasibility of the algorithm, the experiments were conducted on a private dataset provided by a medical institution, and the experimental Dice coefficient reached 0.876 on T2WI images, 0.851 on DWI images, respectively, indicating the existence of an important clinical guidance value for rectal tumor image segmentation algorithms.