Mutually communicated model based on multi‐parametric MRI for automated segmentation and classification of prostate cancer
Kewen Liu,Piqiang Li,Martins Otikovs,Xinzhou Ning,Liyang Xia,Xiangyu Wang,Lian Yang,Feng Pan,Zhi Zhang,Guangyao Wu,Han Xie,Qingjia Bao,Xin Zhou,Chaoyang Liu
DOI: https://doi.org/10.1002/mp.16343
IF: 4.506
2023-03-12
Medical Physics
Abstract:BackGroud Multiparametric magnetic resonance imaging (mp‐MRI) is introduced and established as a noninvasive alternative for Prostate cancer (Pca) detection and characterization. Purpose To develop and evaluate a mutually communicated deep learning segmentation and classification network (MC‐DSCN) based on mp‐MRI for prostate segmentation and prostate cancer (PCa) diagnosis. Methods The proposed MC‐DSCN can transfer mutual information between segmentation and classification components and facilitate each other in a bootstrapping way. For classification task, the MC‐DSCN can transfer the masks produced by the coarse segmentation component to the classification component to exclude irrelevant regions and facilitate classification. For segmentation task, this model can transfer the high‐quality localization information learned by the classification component to the fine segmentation component to mitigate the impact of inaccurate localization on segmentation results. Consecutive MRI exams of patients were retrospectively collected from two medical centers (referred to as center A and B) . Two experienced radiologists segmented the prostate regions, and the ground truth of the classification refers to the prostate biopsy results. MC‐DSCN was designed, trained, and validated using different combinations of distinct MRI sequences as input (e.g., T2‐weighted and apparent diffusion coefficient) and the effect of different architectures on the network's performance was tested and discussed. Data from center A was used for training, validation, and internal testing, while another center's data was used for external testing. The statistical analysis is performed to evaluate the performance of the MC‐DSCN. The DeLong test and paired t‐test were used to assess the performance of classification and segmentation, respectively. Results In total, 134 patients were included. The proposed MC‐DSCN outperforms the networks that were designed solely for segmentation or classification. Regarding the segmentation task, the classification localization information helped to improve the IOU in center A: from 84.5% to 87.8% (p < 0.01) and in center B: from 83.8% to 87.1% (p < 0.01), while the AUC of PCa classification was improved in center A: from 0.946 to 0.991 (p < 0.02) and in center B: from 0.926 to 0.955 (p < 0.01) as a result of the additional information provided by the prostate segmentation. Conclusion The proposed architecture could effectively transfer mutual information between segmentation and classification components and facilitate each other in a bootstrapping way, thus outperforming the networks designed to perform only one task. This article is protected by copyright. All rights reserved
radiology, nuclear medicine & medical imaging