Segmentation versus detection: Development and evaluation of deep learning models for prostate imaging reporting and data system lesions localisation on Bi‐parametric prostate magnetic resonance imaging

Abstract

AbstractAutomated prostate cancer detection in magnetic resonance imaging (MRI) scans is of significant importance for cancer patient management. Most existing computer‐aided diagnosis systems adopt segmentation methods while object detection approaches recently show promising results. The authors have (1) carefully compared performances of most‐developed segmentation and object detection methods in localising prostate imaging reporting and data system (PIRADS)‐labelled prostate lesions on MRI scans; (2) proposed an additional customised set of lesion‐level localisation sensitivity and precision; (3) proposed efficient ways to ensemble the segmentation and object detection methods for improved performances. The ground‐truth (GT) perspective lesion‐level sensitivity and prediction‐perspective lesion‐level precision are reported, to quantify the ratios of true positive voxels being detected by algorithms over the number of voxels in the GT labelled regions and predicted regions. The two networks are trained independently on 549 clinical patients data with PIRADS‐V2 as GT labels, and tested on 161 internal and 100 external MRI scans. At the lesion level, nnDetection outperforms nnUNet for detecting both PIRADS ≥ 3 and PIRADS ≥ 4 lesions in majority cases. For example, at the average false positive prediction per patient being 3, nnDetection achieves a greater Intersection‐of‐Union (IoU)‐based sensitivity than nnUNet for detecting PIRADS ≥ 3 lesions, being 80.78% ± 1.50% versus 60.40% ± 1.64% (p < 0.01). At the voxel level, nnUnet is in general superior or comparable to nnDetection. The proposed ensemble methods achieve improved or comparable lesion‐level accuracy, in all tested clinical scenarios. For example, at 3 false positives, the lesion‐wise ensemble method achieves 82.24% ± 1.43% sensitivity versus 80.78% ± 1.50% (nnDetection) and 60.40% ± 1.64% (nnUNet) for detecting PIRADS ≥ 3 lesions. Consistent conclusions are also drawn from results on the external data set.