Prostate Cancer Detection And PI-RADS Scoring Based On Radiomics And Deep Learning

Prostate cancer（PCa）is the most common cancer in men all over the world.Timely and accurate diagnosis is essential to the health and survival of patients.Conventionally,for PCa diagnosis,transrectal ultrasound-guided prostate biopsy（TRUS-biopsy）represents a standard diagnostic pathway for men with an elevated prostate-specific antigen（PSA）,which also results in unnecessary over biopsy.Prostate Imaging Reporting and Data System（PI-RADS）is the recommended guideline for PCa diagnosis with multiparametric magnetic resonance images（mp MRI）.As a noninvasive imaging technique,it can not only effective reduce the chance of unnecessary biopsy,but can also improving the accuracy of biopsy for those really need,preventing repetitive biopsy.However,accurate interpretation of the mp MRI depends highly on the expertise of radiologists.Thus,great variance exists in the readings of different radiologists,which is one of the biggest difficulties in application of PI-RADS.Therefore,it is significant to use artificial intelligence to help less experienced junior radiologists to evaluate PI-RADS scores accurately and eliminate the inconsistency incurred by subjectivity.Firstly,we built radiomics models for PI-RADS score classification.While most previous radiomics studies focused on differentiating malignant/benign prostate lesions from mp MRI,radiologists depend on PI-RADS scores to make diagnostic plans.Our radiomics models directly classified lesions according to their PI-RADS scores,high grade for PI-RADS≥3 and low grade for the others.First-order and shape features were extracted from regions of interests（ROI）in mp MRI which was annotated by radiologists.After dimension reduction and feature sorting,selected features were used to build classification models based on logistic regression（LR）and support vector machine（SVM）.LR-based and SVM-based models achieved AUC（Area Under receiver operating characteristic Curve）values of 0.918（95%CI[confidence interval]:0.871-0.962）and 0.922（95%CI:0.872-0.964）,respectively,both with a sensitivity of0.723 and specificity of 1.000.Although radiomics can be used to differentiate high grade lesions and low grade lesions,its dependence on manually annotated ROIs limited its application in real clinical settings.Therefore,we constructed two deep learning models,a UNet-Seg for prostate lesion detection and segmentation,and a PI-RADS^AI for PI-RADS scoring,using mp MRI image patches containing segmented lesions as input.PI-RADS^AIintegrated the mp MRI reading process used by subspecialists into the network construction.Besides,considering the ordinal characteristics of PI-RADS scoring task,we incorporated a modified UDM（Unsure Data Model）into the model.Finally,the model trained on a multi-center cohort achieved a Kappa score of 0.761 on PI-RADS four-category classification（score 2-5）,which was significantly higher than those reported previously.In addition,PI-RADS^AI was also compared with subspecialists and general radiologists to evaluate its performance.Though still inferior to subspecialists,PI-RADS^AI had surpassed most of the general radiologists involved and showed its potential to help the general radiologist to improve their performance in PI-RADS scoring.Although the potential of PI-RADS^AI for clinical application had been demonstrated,it required an upstream UNet-Seg for lesion detection and segmentation.Pixel-wise segmentation model would cause some problems such as outputting some little noisy regions.Moreover,these two models cannot be trained jointly and form an end-to-end model.For these reasons,we constructed a lesion-wise detection model.We incorporated clinical prior into FCOS detector,which was used wide in computer vision and modified the some modules to build a so-called o FCOS（ordinal FCOS）model.The proposed o FCOS achieved an AUC value of 0.825 for lesion detection and a Kappa score of 0.656 for PI-RADS classification.In summary,we constructed an integrated pipeline to detect prostate lesions and to predict corresponding PI-RADS scores.The results showed that the model achieved a performance comparable to general radiologists and could potentially be used in clinical settings.