A comparative study of automatic hippocampal segmentation in whole-brain radiotherapy applying DPNU-Net, Mask-RCNN, and nnU-Net models

Abstract

Purpose

This study developed a hippocampal segmentation model that can be used by clinicians by applying the Dual Path Networks U-Net (DPNU-Net), Mask-Region Convolution Neural Networks (Mask-RCNN), and No New U-Net (nnU-Net) algorithms for segmenting the hippocampus to provide a reference for the accurate implementation of hippocampal-avoidance whole-brain radiotherapy (HA-WBRT).

Methods

We retrospectively collected T1-weighted imaging (T1WI) and contrast-enhanced (CE)- T1WI sequence magnetic resonance images of 312 patients with brain metastases, of which 62 served as the test set and 250 as the training set. Manual segmentation was used as the gold standard to compare the differences in the dice similarity coefficient (DSC), relative volume error (RVE), 95% hausdorff distance (95%HD) and average surface distance (ASD) of the DPNU-Net, Mask-RCNN, and nnU-Net models for segmenting the hippocampus in different images.

Results

(1) Compared with manual segmentation, the DPNU-Net, Mask-RCNN, and nnU-Net models segmented the hippocampus based on T1WI with DSCs of 0.819–0.897, RVEs of −3.40% to 1.40%, 95%HDs of 0.813–37.425 mm, ASDs of 0.155–3.907 mm. The best segmentation effect was achieved by using the DPNU-Net model. (2) Based on the CE-T1WI, the DSCs of the DPNU-Net, Mask-RCNN, and nnU-Net models compared with manual segmentation were 0.791–0.879, whereas the RVEs were −1.90% to 2.20%, 95%HDs were 0.915–47.812 mm, and the ASDs were 0.210–5.384 mm. The segmentation effect of the DPNU-Net model was the best. (3) When comparing the DPNU-Net, Mask-RCNN, and nnU-Net models for T1WI and CE-T1WI segmentation, the differences in the DSC, 95%HD, ASD and volumes of the hippocampi segmented by the DPNU-Net model between the two sequences of images were the smallest.

Conclusions

Considering automatic hippocampal segmentation, the DPNU-Net model had a higher accuracy and more stable performance than the Mask-RCNN and nnU-Net models. Thus the DPNU-Net model can be used as a practical method for automatic hippocampal segmentation in the HA-WBRT technique.