Study On Automatic Segmentation And T-stage Diagnosis Of Prostate Cancer Based On Deep Learning On MRI Images

Objective:Prostate cancer(PCa)is an aggressive malignant tumor that occurs in the epithelial cells of prostate,which is one of the most common malignancies in men due to the increasing proportion of adults and aging population resulting in an increasing morbidity and mortality year by year.Early detection and early treatment are the key to improving the prognosis.Magnetic resonance imaging(MRI)images can better display the location of lesions.However,there are still many problems in the current diagnosis and treatment.Firstly,uneven distribution of medical resources leading to high rates of underdiagnosis and misdiagnosis of diseases,and poor prognosis of patients.Secondly,patients were not treated promptly and missed the best treatment period.Thirdly,multidisciplinary collaborative model cannot be popularized.With the rapid development of artificial intelligence in the medical field,especially in big data analysis,images recognition and classification,medical images have gradually transformed from an initial auxiliary diagnostic tool into a core role in personalized precision medicine.To solve the above problems,we used manually annotated datasets to perform anatomical statistical analysis of benign and malignant prostate diseases and T stages of PCa and created automatic segmentation models and automatic T-stages diagnostic models based on deep learning to assist clinicians in diagnosis.Methods:The preoperative MRI images of 542 patients from Shanxi Cancer Hospital from April2018 to June 2020(n=33),Second Affiliate Hospital of Army Medical University from January 2020 to January 2022(n=196),and the public dataset of PROSTATE-x(n=345)were retrospectively collected.Among them,175 patients with PCa have complete clinical information.1.Construction of statistical anatomical modelThe 175 patients with PCa were divided into 4 groups based on clinical information included T1 group(n=25),T2 group(n=80),T3 group(n=36),and T4 group(n=34).Three-dimensional models were reconstructed with manually segmentation of peripheral zone(PZ),transition zone(TZ),central zone(CZ),anterior fibromuscular stroma(AFS),tumors,and their volume,long axis,cross-section and volume ratio were calculated to analyze the differences of anatomical parameters at different T-stages.2.Construction of automatic segmentation modelThe prostatic pre-MRI images of 542 patients were randomly divided into training(n=381),validation(n=53)and test cohort(n=108)according to the ratio of 7:1:2.The manual annotations of prostate zones,tumors,and adjacent structures were regarded as ground truth.And the segmentation performance was quantitatively evaluated with the Dice similarity(DSC),precision,intersection over union(Io U),95%Hausdoff distance(95%HD),and volume difference in test cohort.3.Construction of automatic T-stage diagnosis modelThe pre-MRI images of 175 PCa patients were randomly divided into training(n=127),validation(n=18),and test cohort(n=30)by the ratio of 7:1:2 for training in automated T-stage diagnosis.Diagnosis performances were evaluated using precision,specificity,F1-score,accuracy,and area under the receiver operating characteristic curve(AUC).Results:1.Statistical analysis of anatomical parameters.Tumor’s major axis,tumor’s cross-sectional area,tumor’s volume,and the volume ratio of tumor and prostate can be used to diagnose T-stage diagnosis of PCa,with cutoff values of 2.30 mm,1.30 cm~2,0.90 cm~3,and0.03 for T1-T2 stage,2.80 mm,4.10 cm~2,8.29 cm~3,and 2.80 for T2-T3 stage,and 4.17 mm,3.84 mm~2,3.84 cm~3,and 0.22 for T3-T4 stage,respectively.2.Evaluation of segmentation models.The VA-Unet model showed better segment performance of PZ,TZ,CZ,AFS,tumors,urethra,seminal vesicle,bladder,and rectum with DSC of 0.816,0.903,0.501,0.586,0.583,0.613,0.738,0.926,0.848 and 95%HD of16.10mm,12.61mm,17.37mm,21.77mm,40.29 mm,11.95 mm,25.93 mm,34.67 mm,and7.60 mm,respectively.Except for urethra,the volume of other structures are good consistent with the manual annotation,and the difference is basically within the 95%consistent range.3.Evaluation of T-stage diagnosis model.Dense-Net,Res-Net,Vi T,and Swin T based on the predict results of VA-Unet,performance better than that on the whole MRI images with AUCs of 0.648,0.491,0.787,and 0.796,and accuracy of 0.587,0.600,0.700,and 0.733.Conclusion:The anatomical parameters based on MRI images can be used for the differential diagnosis of T-stage diagnosis of PCa.And the proposed VAU-Swin T model can achieve accurate and automatic segmentation of prostate zones,tumor,and adjacent structures and automatic T-stage diagnosis,which effectively improve the diagnosis efficiency of PCa,reduce the doctors’working hours and workload,save the medical sources,and promote the development of prostate MRI image diagnosis technology.