AI-assisted biparametric MRI surveillance of prostate cancer: feasibility study

Abstract

Objectives: To evaluate the feasibility of automatic longitudinal analysis of consecutive biparametric MRI (bpMRI) scans to detect clinically significant (cs) prostate cancer (PCa).

Methods: This retrospective study included a multi-center dataset of 1513 patients who underwent bpMRI (T2 + DWI) between 2014 and 2020, of whom 73 patients underwent at least two consecutive bpMRI scans and repeat biopsies. A deep learning PCa detection model was developed to produce a heatmap of all PIRADS ≥ 2 lesions across prior and current studies. The heatmaps for each patient's prior and current examination were used to extract differential volumetric and likelihood features reflecting explainable changes between examinations. A machine learning classifier was trained to predict from these features csPCa (ISUP > 1) at the current examination according to biopsy. A classifier trained on the current study only was developed for comparison. An extended classifier was developed to incorporate clinical parameters (PSA, PSA density, and age). The cross-validated diagnostic accuracies were compared using ROC analysis. The diagnostic performance of the best model was compared to the radiologist scores.

Results: The model including prior and current study (AUC 0.81, CI: 0.69, 0.91) resulted in a higher (p = 0.04) diagnostic accuracy than the current only model (AUC 0.73, CI: 0.61, 0.84). Adding clinical variables further improved diagnostic performance (AUC 0.86, CI: 0.77, 0.93). The diagnostic performance of the surveillance AI model was significantly better (p = 0.02) than of radiologists (AUC 0.69, CI: 0.54, 0.81).

Conclusions: Our proposed AI-assisted surveillance of prostate MRI can pick up explainable, diagnostically relevant changes with promising diagnostic accuracy.

Key points: • Sequential prostate MRI scans can be automatically evaluated using a hybrid deep learning and machine learning approach. • The diagnostic accuracy of our csPCa detection AI model improved by including clinical parameters.

Keywords: Deep learning; Magnetic resonance imaging; Prostatic neoplasms.

Fig. 1

Overview of the division of the dataset

Fig. 2

A visualization of the pipeline used to predict the likelihood of csPCa using sequential data (D_seq). First, the two scans of a single patient are separately processed by the detection U-Net to generate two independent heatmaps for detected lesions (HM1, HM2). The U-Net was previously trained using dataset D_det. Features for the likelihood and volume are then extracted from each of the heatmaps and used by the multi-scan classifier to predict the likelihood of csPCa at the time of the current examination

Fig. 3

Examination-level ROC curves for the detection of csPCa at follow-up for each of the AI classifiers described in “Experiments.” Transparent areas behind each of the curves indicate bootstrapped 95% confidence intervals. The ROC for the radiologist assigned PIRADS scores is shown as the dotted line. Cross markers for the AI models indicate the sensitivity and specificity at the optimal cutoff point determined using Youden’s J statistic

Fig. 4

A 70-year-old man with a transitional zone lesion (delineated by the purple outline in each image) that was correctly diagnosed as non-csPCa by the surveillance AI. The initial biopsy at the prior scan was negative for PCa, and the targeted repeat biopsy at the current examination showed non-csPCa (Gleason 3 + 3 = 6). The figure shows T2-weighted images (a, b) and ADC maps (c, d), and corresponding detection heatmaps for the prior and current examination (e, f). The detection heatmap of the second examination showed a mostly unchanged volume and likelihood score. The radiologist assigned a score of PIRADS 4

Fig. 5

T2-weighted images (a, b) and ADC maps (c, d), and corresponding detection heatmaps (e, f) showing progression in a detected transitional zone lesion in a 66-year-old man on AS. The heatmaps show that a lesion was detected by the U-Net in both scans, and an increase in likelihood score and volume was recorded over a 4-year follow-up. The radiologist reported no PCa localization in the scan and assigned a score of PIRADS 1. A systematic repeat biopsy at the current examination revealed a Gleason score of 3 + 4 = 7. The surveillance AI correctly classified this patient as csPCa based on both scans