Multi-organ segmentation of abdominal structures from non-contrast and contrast enhanced CT images

Abstract

Manually delineating upper abdominal organs at risk (OARs) is a time-consuming task. To develop a deep-learning-based tool for accurate and robust auto-segmentation of these OARs, forty pancreatic cancer patients with contrast-enhanced breath-hold computed tomographic (CT) images were selected. We trained a three-dimensional (3D) U-Net ensemble that automatically segments all organ contours concurrently with the self-configuring nnU-Net framework. Our tool’s performance was assessed on a held-out test set of 30 patients quantitatively. Five radiation oncologists from three different institutions assessed the performance of the tool using a 5-point Likert scale on an additional 75 randomly selected test patients. The mean (± std. dev.) Dice similarity coefficient values between the automatic segmentation and the ground truth on contrast-enhanced CT images were 0.80 ± 0.08, 0.89 ± 0.05, 0.90 ± 0.06, 0.92 ± 0.03, 0.96 ± 0.01, 0.97 ± 0.01, 0.96 ± 0.01, and 0.96 ± 0.01 for the duodenum, small bowel, large bowel, stomach, liver, spleen, right kidney, and left kidney, respectively. 89.3% (contrast-enhanced) and 85.3% (non-contrast-enhanced) of duodenum contours were scored as a 3 or above, which required only minor edits. More than 90% of the other organs’ contours were scored as a 3 or above. Our tool achieved a high level of clinical acceptability with a small training dataset and provides accurate contours for treatment planning.

Automatic end‐to‐end VMAT treatment planning for rectal cancers

“…For auto‐segmentation, previous studies have used more than 200 patient scans to train for OAR segmentations in rectal cancers on CT scans 7,15 . This study showed that high clinical acceptability of OAR contours can be obtained with a small but consistent and well‐curated training dataset, consistent with a recent finding using nnU‐Net 13 . The automated contouring of large and small bowel can facilitate clinicians to employ separate dose objectives for different bowel structures, which often are too time‐consuming to contour without the aid of auto‐segmentation.…”

“…Therefore, each OAR structure had 18 contours from 18 patients. Though these patient numbers were relatively low, a previous study has shown that clinically acceptable predictions can be achieved with consistent data curation using low numbers of patient scans 13 . The slice thicknesses of the CT scans were either 2, 2.5, or 3 mm.…”

Dose Escalation for Pancreas SBRT: Potential and Limitations of using Daily Online Adaptive Radiation Therapy and an Iterative Isotoxicity Automated Planning Approach

“…Currently, the estimated time to generate a dose-escalated adaptive plan using the automation algorithms is less than 30 minutes (8 minutes for our autocontouring algorithm from our previous study 15 and 14.8 ± 6.5 minutes [mean ± 1σ] for the dose-escalated daily adaptive planning algorithm). The process time for the autoplanning algorithm might be further reduced if we can quickly predict the achievable prescription dose with given patient images 19 , 20 , 21 , 22 and use the predicted value as the target dose constraints instead of using the iterative algorithm described in Figure 1 A.…”

“…We also generated the contours for 9 OARs (stomach, small bowel, large bowel, duodenum, spinal cord, left and right kidneys, liver, and skin) in the abdominal region. The OAR contours were automatically generated using an in-house deep-learning-based auto-contouring system 15 and were subsequently reviewed and edited by experienced medical physicists. The skin contour was defined to be the 2-cm thick internal ring from the surface of the body.…”

A new architecture combining convolutional and transformer‐based networks for automatic 3D multi‐organ segmentation on CT images

“…Recently deep learning algorithms, which rely primarily on fully convolutional neural networks (CNN)based U-net architectures [16][17][18][19][20][21][22][23] have been applied to the problem of organ segmentation for radiation treatment planning. [24][25][26] The U-Net is a popular architecture and comprises an encoder and decoder, where the encoder progressively reduces the resolution of CT scans to generate conceptual features across multiple scales.…”