Augmenting AI with Automated Segmentation of Report Findings Applied to Prostate Cancer Detection in Biparametric MRI

J. Bosma, A. Saha, M. Hosseinzadeh and H. Huisman

Master thesis 2021.


Prostate MRI interpreted by expert radiologists provides the best non-invasive diagnosis of clinically significant prostate cancer (csPCa), but is a limited resource. Deep learning has the potential to assist, but requires expensive expert annotations. We developed an automatic labelling procedure guided by radiology reports from clinical routine, capable of generating high quality voxel-level annotations. First, we parse the radiology report to extract the number of clinically significant findings (nsig), and then generate annotation by keeping the nsig most confident lesion candidates from a csPCa segmentation model. We included 7,756 prostate bpMRI studies (axial T2, high b-value and ADC scans), of which 3,052 were manually annotated and 4,704 were automatically annotated, resulting in the largest prostate MRI dataset reported in literature. We evaluated the automatic annotation procedure using the manual annotations: our score extraction correctly identified nsig for 99.3% of the visits, our prostate cancer segmentation model correctly localised 83.7 +- 3.0% of the lesions, and the automatic annotations of correctly localised lesions have good spatial congruence, with Dice similarity coefficients of 0.71+-0.15. Augmenting the training set with automatically labelled visits significantly improved prostate cancer detection performance, as evaluated on 296 visits from an independent, external, centre with ground truth provided by MR-guided and TRUS-guided biopsies, or radical prostatectomy when available. Patient-based diagnostic AUROC increased from 86.6 +- 1.8% to 88.7 +- 1.1% (P = 0.047) and lesion-based diagnostic pAUC increased from 1.940+-0.082 to 2.050+-0.031 (P = 0.016), with mean+-std. over 25 independent runs.