Abstract / Summary
Abstract Objectives To develop and test deep-learning models which detect extraprostatic extension (EPE) on biparametric (bpMRI) and multiparametric MRI (mpMRI) of the prostate in comparison to radiologists. Methods Consecutive patients at a large healthcare enterprise who underwent prostate MRI (2015 to 2023) with subsequent radical prostatectomy within 1 year were included. The dataset was divided into training/validation/test sets. Transfer learning models, composed of two multi-branch 3D convolutional neural networks based on a 3D ResNet-50 backbone, were trained on bpMRI (AI bp ) with an input of axial T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) and on mpMRI (AI mp ) with an input of axial T2WI, axial DWI, and axial post-contrast imaging. No prostate gland or tumor segmentations were performed. A logistic regression model with prostate specific antigen density was also trained with the AI mp model (AI mpPSAD ). Three fellowship-trained abdominal radiologists evaluated the test set for EPE on a 1 to 5 scoring system based on capsular appearance, on bpMRI and mpMRI for each study. Areas under the receiver operating characteristic curves (AUC) were obtained and compared with Delong test. Diagnostic statistics were also obtained. Results 1,232 prostate MRIs were included (1003/113/116 in the training/validation/test sets). AUC for models were as follows: AI bp (0.71), AI mp (0.72), and AI mpPSAD (0.73). While holding sensitivity constant at the Youden index for AI bp (0.77), specificity and accuracy of the AI bp , AI mp , and AI mpPSAD were 0.64, 0.69, 0.71 and 0.71, 0.73, 0.74, respectively. The ranges of AUC for readers were: bpMRI (0.66–0.71) and mpMRI (0.69–0.74). The ranges of accuracy of the readers on bpMRI was 0.60–0.68 and on mpMRI was 0.65–0.72. Conclusion A segmentation-free transfer learning model performs comparably to fellowship-trained radiologists on the detection of EPE on both bpMRI and mpMRI. There was a trend of increasing accuracy with the addition of post contrast sequences for all models and readers.