Characterizing the Performance of a Deep-Learning Segmentation Algorithm for Outlining Dominant Intraprostatic Lesions on Prostate Multi-Parametric MRI

Abstract

<h3>Purpose/Objective(s)</h3> Accurately segmenting dominant intraprostatic lesions (DILs) from prostate multi-parametric magnetic resonance images (mpMRI) may have significant clinical implications. However, DIL segmentation can be time-consuming, and subject to considerable inter-observer variability. We hypothesize that an artificial-intelligence (AI) deep-learning segmentation algorithm can provide reliable DIL contours that are not significantly different from those provided by radiation oncologists (ROs). <h3>Materials/Methods</h3> We conducted a single-institutional retrospective study of 81 patients who underwent 3.0 Tesla, mpMRI (GE; B-value >= 1400; endorectal coil), followed by radiation therapy (RT) between 2010-2014. An expert radiologist, blinded to all clinical data, retrospectively identified all PI-RADS 3-5 lesions. A RO contoured the prostate, peripheral zone (PZ), and DILs, to serve as reference labels. For deep learning, we used a 3D-nnUNet (Isensee, Nat Med, 2020), with apparent diffusion coefficient and diffusion-weighted images (DWI) as inputs. We did not use T2-weighted images, due to artifacts from recent biopsies and complex deformations from DWI images. The cohort was partitioned into training (n = 40) and testing (n = 41) sets. Testing patients were identified by enrollment on a prospective registry. We performed 5-fold cross-validation for the training set, and then applied the AI algorithm onto the test set. We evaluated the diagnostic performance (sensitivity, Dice coefficient) of the AI segmentations against reference labels. To test whether AI-contours were similar to RO-provided contours, 5 ROs independently provided contours for the subgroup of patients with PI-RADS4/5 lesions. We tested whether there was a significant difference in Dice coefficients among the 6 contour sets (AI + 5 RO) using the non-parametric Friedman test. <h3>Results</h3> For the testing set, median Dice coefficient, per-patient sensitivity, and per-lesion sensitivities were 0.68, 90.2%, and 77.8%, respectively. Respective values for the 35-patient subgroup with PI-RADS4-5 DILs were 0.75, 94.3%, and 82.6%. The Pearson's correlation between AI and reference DIL volumes was 0.94. Among the 35 patients with PI-RADS4-5 lesions, there was no significant difference in Dice coefficients among the 6 contour sets (p = 0.79). <h3>Conclusion</h3> Deep-learning may provide reliable contours of PI-RADS 4-5 DILs that are similar to those provided by expert radiation oncologists. Further prospective evaluation is warranted.