N Dang*, Y Zhang, A Amjad, J Ding, C Sarosiek, X Li, Medical College of Wisconsin, Milwaukee, WI

• 
  N Dang

TU-D1030-IePD-F2-4 (Tuesday, 7/12/2022) 10:30 AM - 11:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: Deep learning auto-segmentation (DLAS), although very promising, still has limited success particularly for complex structures such as those in abdomen. Consequently, the contours from DLAS need to be visually evaluated and edited, if necessary. This manual evaluation process can be time consuming, especially in MR-guided adaptive radiotherapy (MRgART) where fast and accurate auto-segmentation on daily image is essential. Building up on our previous studies, this work aims to develop a deep learning based automatic contour quality assurance (ACQA) classification model to quickly evaluate auto-segmented abdominal contours on MRI for MRgART.

Methods: The ACQA model based on a convolutional neural network (U-net) was trained and tested using 3000 image slices along with auto-segmented contours of abdominal MRIs acquired for 65 patients. The training and testing sets included 80% and 20% of the data, respectively. Quality of auto-segmented contours was rated as acceptable and unacceptable with minor or major edit using a classification model developed in a separate study. The performance of the ACQA model was evaluated using accuracy and area under a receiver operating characteristic (ROC) curve (AUC).

Results: Accuracies of the obtained ACQA model were 0.83, 0.89, 0.88, 0.91, and 0.90 for stomach, duodenum, pancreas, small, and large bowels, respectively. The highest accuracy was obtained for the acceptable contour category (>95%), and the lowest was for the minor edit category (83%). In the best classification performance case (small bowel), the AUCs were 0.993, 0.979 and 0.994 for acceptable, minor edit and major edit categories, respectively.

Conclusion: The newly developed deep learning based ACQA model can quickly and automatically classify quality of auto-segmentation of complex abdominal organs. With further development using datasets with more variety, this model can be integrated into an auto-segmentation workflow, facilitating efficient and accurate segmentation in routine RT planning or MRgART.

Keywords

Quality Assurance, Segmentation, MRI

Taxonomy

IM/TH- image Segmentation: General (Most aspects)

Contact Email