Automatic segmentation of magnetic resonance images for high-dose-rate cervical cancer brachytherapy using deep...

Magnetic resonance (MR) imaging is the gold standard in image-guided brachytherapy (IGBT) due to its superior soft-tissue contrast for target and organs-at-risk (OARs) delineation. Accurate and fast segmentation of MR images are very important for high-quality IGBT treatment planning. The purpose of this work is to implement and evaluate deep learning (DL) models for the automatic segmentation of targets and OARs in MR image-based high-dose-rate (HDR) brachytherapy for cervical cancer.

A 2D DL model using residual neural network architecture (ResNet50) was developed to contour the targets (gross tumor volume (GTV), high-risk clinical target volume (HR CTV), and intermediate-risk clinical target volume (IR CTV)) and OARs (bladder, rectum, sigmoid, and small intestine) automatically on axial MR slices of HDR brachytherapy patients. Furthermore, two additional 2D DL models using sagittal and coronal images were also developed. A 2.5D model was generated by combining the outputs from axial, sagittal, and coronal DL models. Similarly, a 2D and 2.5D DL models were also generated for the inception residual neural network (InceptionResNetv2 (InRN)) architecture. The geometric (Dice similarity coefficient (DSCs) and 95th percentile of Hausdorff distance (HD)) and dosimetric accuracy of 2D (axial only) and 2.5D (axial + sagittal + coronal) DL model generated contours were calculated and compared.

The mean (range) DSCs of ResNet50 across all contours were 0.674 (0.05-0.96) and 0.715 (0.26-0.96) for the 2D and 2.5D models, respectively. For InRN, these were 0.676 (0.11-0.96) and 0.723 (0.35-0.97) for the 2D and 2.5D models, respectively. The mean HD of ResNet50 across all contours was 15.6 mm (1.8-69 mm) and 12.1 mm (1.7-44 mm) for the 2D and 2.5D models, respectively. The similar results for InRN were 15.4 mm (2-68 mm) and 10.3 mm (2.7-39 mm) for the 2D and 2.5D models, respectively. The dosimetric parameters (D90) of GTV and HR CTV for manually contoured plans matched better with the 2.5D model (p > 0.6) and the results from the 2D model were slightly lower (p < 0.08). On the other hand, the IR CTV doses (D90) for all of the models were slightly lower (2D: -1.3 to -1.5 Gy and 2.5D: -0.5 to -0.6 Gy) and the differences were statistically significant for the 2D model (2D: p < 0.000002 and 2.5D: p > 0.06). In case of OARs, the 2.5D model segmentations resulted in closer dosimetry than 2D models (2D: p = 0.07-0.91 and 2.5D: p = 0.16-1.0).

The 2.5D DL models outperformed their respective 2D models for the automatic contouring of targets and OARs in MR image-based HDR brachytherapy for cervical cancer. The InceptionResNetv2 model performed slightly better than ResNet50.