Intraprostatic Tumor Segmentation on PSMA PET Images in Patients with Primary Prostate Cancer with a Convolutional Neural Network

Abstract

Accurate delineation of the intraprostatic gross tumor volume (GTV) is a prerequisite for treatment approaches in patients with primary prostate cancer (PCa). Prostate-specific membrane antigen PET (PSMA PET) may outperform MRI in GTV detection. However, visual GTV delineation underlies interobserver heterogeneity and is time consuming. The aim of this study was to develop a convolutional neural network (CNN) for automated segmentation of intraprostatic tumor (GTV-CNN) in PSMA PET. Methods: The CNN (3D U-Net) was trained on the ⁶⁸Ga-PSMA PET images of 152 patients from 2 different institutions, and the training labels were generated manually using a validated technique. The CNN was tested on 2 independent internal (cohort 1: ⁶⁸Ga-PSMA PET, n = 18 and cohort 2: ¹⁸F-PSMA PET, n = 19) and 1 external (cohort 3: ⁶⁸Ga-PSMA PET, n = 20) test datasets. Accordance between manual contours and GTV-CNN was assessed with the Dice-Sørensen coefficient (DSC). Sensitivity and specificity were calculated for the 2 internal test datasets (cohort 1: n = 18, cohort 2: n = 11) using whole-mount histology. Results: The median DSCs for cohorts 1-3 were 0.84 (range: 0.32-0.95), 0.81 (range: 0.28-0.93), and 0.83 (range: 0.32-0.93), respectively. Sensitivities and specificities for the GTV-CNN were comparable with manual expert contours: 0.98 and 0.76 (cohort 1) and 1 and 0.57 (cohort 2), respectively. Computation time was around 6 s for a standard dataset. Conclusion: The application of a CNN for automated contouring of intraprostatic GTV in ⁶⁸Ga-PSMA and ¹⁸F-PSMA PET images resulted in a high concordance with expert contours and in high sensitivities and specificities in comparison with histology as a reference. This robust, accurate and fast technique may be implemented for treatment concepts in primary prostate cancer. The trained model and the study's source code are available in an open source repository.

Keywords: PSMA-PET; convolutional neuronal networks; histopathology; prostate cancer; segmentation.

FIGURE 1.

Visualization of the training and evaluation curves. Shown are training and evaluation results as dice loss and DSC (A) and HD and ASSD (B).

FIGURE 2.

Histology reference projected on ⁶⁸Ga-PSMA-11 (upper row) and ¹⁸F-PSMA-1007 (lower row). (A and C) Hematoxylin and eosin whole-mount prostate slide with marked PCa lesion. (B and D) Axial PET image (image windowing: SUVmin-max = 0–5). Blue contour = prostate; green contour = histology reference; yellow contour = GTV-Exp; yellow contour = GTV-CNN.

FIGURE 3.

Specificity and sensitivity of GTV-CNN, GTV-Exp, and GTV-30% based on comparison with histology reference. (Upper row) Cohort 1 (⁶⁸Ga-PSMA-11 PET). (Lower row) Cohort 2 (¹⁸F-PSMA-1007 PET). Box plots are presented. Pairwise comparison was performed with Wilcoxon signed-rank test. n.s. = not significant. *P = 0.05–0.01.