Ensemble learning and tensor regularization for cone‐beam computed tomography‐based pelvic organ segmentation

Abstract

Purpose Cone‐beam computed tomography (CBCT) is a widely accessible low‐dose imaging approach compatible with on‐table patient anatomy observation for radiotherapy. However, its use in comprehensive anatomy monitoring is hindered by low contrast and low signal‐to‐noise ratio and a large presence of artifacts, resulting in difficulty in identifying organ and structure boundaries either manually or automatically. In this study, we propose and develop an ensemble deep‐learning model to segment post‐prostatectomy organs automatically. Methods We utilize the ensemble logic in various modules during the segmentation process to alleviate the impact of low image quality of CBCT. Specifically, (1) semantic attention was obtained from an ensemble 2.5D You‐only‐look‐once detector to consistently define regions of interest, (2) multiple view‐specific two‐stream 2.5D segmentation networks were developed, using auxiliary high‐quality CT data to aid CBCT segmentation, and (3) a novel tensor‐regularized ensemble scheme was proposed to aggregate the estimates from multiple views and regularize the spatial integrity of the final segmentation. Results A cross‐validation study achieved Dice similarity coefficient and mean surface distance of 0.779 ±$\pm$ 0.069 and 2.895 ±$\pm$ 1.496 mm for the rectum, and 0.915 ±$\pm$ 0.055 and 1.675 ±$\pm$ 1.311 mm for the bladder. Conclusions The proposed ensemble scheme manages to enhance the geometric integrity and robustness of the contours derived from CBCT with light network components. The tensor regularization approach generates organ results conforming to anatomy and physiology, without compromising typical quantitative performance in Dice similarity coefficient and mean surface distance, to support further clinical interpretation and decision making.

Technical note: Air bubble‐induced performance degradation in automatic rectum segmentation from cone‐beam CT

“…All images were manually contoured by radiation oncologists. Automatic segmentation of the rectum was performed for each of the axial, sagittal, and coronal views using a 2.5D network with residual blocks 10 …”

“…While automatic segmentation using deep networks has demonstrated promise, it has shown inferior performance based on CBCT than fan‐beam CT with similar networks 1,8,9 . In addition, we have observed in a recent study with 85 CBCT cases that rectum segmentation from postprostatectomy patients has been generally more challenging than from patients without surgery, even with comparable rectum morphology statistics 10 . We hypothesize that this may be related to the higher proportion of postsurgical patients who experience abnormal bowel movements and constipation, resulting in much more frequent and severe air bubble presence in the rectum 11 .…”

“…1,8,9 In addition, we have observed in a recent study with 85 CBCT cases that rectum segmentation from postprostatectomy patients has been generally more challenging than from patients without surgery, even with comparable rectum morphology statistics. 10 We hypothesize that this may be related to the higher proportion of postsurgical patients who experience abnormal bowel movements and constipation, resulting in much more frequent and severe air bubble presence in the rectum. 11 In this work, we systematically examine the relation between air bubble severity and the task-specific endpoint of deep-learning-based automatic rectum segmentation.…”

Evaluation of generalization ability for deep learning‐based auto‐segmentation accuracy in limited field of view CBCT of male pelvic region

“…[10][11][12] Researchers have conducted segmentation of soft tissue organs on CT and/or CBCT using neural encoding/decoding based on the CNN architecture, thereby exploiting supervised training. [13][14][15][16][17][18][19][20][21] The main requirement for this type of approach is anatomical correspondence between the input image (CT or CBCT), and the ground truth reference label. [13][14][15][16][17][18][19][20][21] Although CNNbased segmentation works well for full field of view (FOV) CBCT, [13][14][15][16][17]20,21 it is unknown whether CNNbased segmentation using limited FOV CBCT is possible.…”

Improving performance of deep learning models using 3.5D U-Net via majority voting for tooth segmentation on cone beam computed tomography

“…Several U-Nets including 2D U-Net 20,21 , 2.5D U-Net 22 , and 3D U-Net 23 have been proposed for CBCT segmentation. A variant of 2.5D U-Net using majority voting of 2D U-Nets trained by 3 orthogonal imaging planes has been shown to outperform any single U-Net for maxillary and mandibular bony structure segmentation on CBCT 24 .…”