Accuracy Validation of an Automated Method for Prostate Segmentation in Magnetic Resonance Imaging

Abstract

Three dimensional (3D) manual segmentation of the prostate on magnetic resonance imaging (MRI) is a laborious and time-consuming task that is subject to inter-observer variability. In this study, we developed a fully automatic segmentation algorithm for T2-weighted endorectal prostate MRI and evaluated its accuracy within different regions of interest using a set of complementary error metrics. Our dataset contained 42 T2-weighted endorectal MRI from prostate cancer patients. The prostate was manually segmented by one observer on all of the images and by two other observers on a subset of 10 images. The algorithm first coarsely localizes the prostate in the image using a template matching technique. Then, it defines the prostate surface using learned shape and appearance information from a set of training images. To evaluate the algorithm, we assessed the error metric values in the context of measured inter-observer variability and compared performance to that of our previously published semi-automatic approach. The automatic algorithm needed an average execution time of ∼60 s to segment the prostate in 3D. When compared to a single-observer reference standard, the automatic algorithm has an average mean absolute distance of 2.8 mm, Dice similarity coefficient of 82%, recall of 82%, precision of 84%, and volume difference of 0.5 cm³ in the mid-gland. Concordant with other studies, accuracy was highest in the mid-gland and lower in the apex and base. Loss of accuracy with respect to the semi-automatic algorithm was less than the measured inter-observer variability in manual segmentation for the same task.

Keywords: 3D segmentation; Automatic segmentation; Endorectal receive coil; Image segmentation; Magnetic resonance imaging; Validation.

Fig. 1

Automatic coarse localization of the prostate. The dashed line shows the estimated tangent line to the rectal wall. The dashed curve shows the estimated bladder border. The solid line polygon is the template used to select the center points for apex, mid-gland, and base. The prostate boundary based on manual segmentation has been overlaid with a d otted line for reference. AP and IS are, respectively, anterioposterior and inferior-superior dimensions of the prostate measured during routine clinical ultrasound imaging. The three indicated points on the template define the three estimated center points for the prostate

Fig. 2

Elements used to calculate the DSC, recall, and precision validation metrics. X and Y are the two shapes, with Y taken as the reference shape. FP false positive, TP true positive, FN false negative

Fig. 3

Qualitative results of automatic, semi-automatic, and manual segmentations for three sample prostates. Each row shows the results at three 2D cross sections of one prostate: the left one at apex subregion, the middle one at mid-gland subregion, and the right one at base subregion. The automatic algorithm’s segmentations are shown with solid magenta contours, the semi-automatic algorithm’s segmentations are showed with dashed blue contours, and the manual segmentations are shown with dotted green contours (color figure online)

Fig. 4

Accuracy of the computer-based segmentations versus inter-operator variability of manual segmentation. The average accuracy of one set of 10 automatic and nine sets of 10 semi-automatic segmentations in comparison with three manual reference segmentations in terms of (a) MAD, (b) DSC, (c) recall, (d) precision, and (e) ΔV. The dashed line segments show the observed range of each metric at each ROI in pairwise comparison between three manual segmentations. For ΔV, the ranges are based on the absolute value of ΔV due to lack of reference in comparison of two manual segmentations. The error bars show one standard deviation. The significant differences detected between semi-automatic and automatic segmentation at different ROIs have been indicated on the graphs with an asterisk (p value < 0.05)

Fig. 5

Accuracy of the computer-based segmentations versus inter-operator variability of manual segmentation. The average accuracy of one set of 10 automatic and nine sets of 10 semi-automatic segmentations in comparison with STAPLE reference segmentation in terms of (a) MAD, (b) DSC, (c) recall, (d) precision, and (e) ΔV. The dashed line segments show the observed range of each metric at each ROI in comparison between three manual segmentations and STAPLE reference. The error bars show one standard deviation. The significant differences detected between semi-automatic and automatic segmentation at different ROIs have been indicated on the graphs (p value < 0.05)