Click here to

Session: AI-Based Auto-Segmentation and Auto-Contouring - I [Return to Session]

Towards Artificial Intelligence and Clinician Integrated Systems (AICIS): Interactive Contour Revision with Deep Boundary Net

T Bai*, A Balagopal, M Dohopolski, H Morgan, R McBeth, J Tan, D Nguyen, S Jiang, UT Southwestern Medical Center, Dallas, TX

Presentations

TU-E-TRACK 6-7 (Tuesday, 7/27/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: Accurate organ contours are required to ensure the quality of a radiotherapy treatment plan. However, contours generated by current auto-segmentation tools are not perfect, will never be, and may require significant revisions which can be tedious and time consuming. This work proposes an artificial intelligence and clinician integrated systems (AICIS) that can interact and assist clinicians to quickly revise the contours.

Methods: The pipeline starts with a convolutional neural network (CNN)-based auto-segmentation model, which generates an initial contour. After review by the clinician, if it is not acceptable, the clinician will click on the boundary that exhibits large errors. This boundary click will be converted into a Gaussian point and then fed into the network to guide an interactive model revising the contour. This interaction process can be iterated multiple times. The input of our interactive model consists of three channels: the CT image, the current segmentation result, and the Gaussian point. The training dataset is built upon the existing dataset for auto-segmentation model training by simulating clinician’s click. In this work, we reasonably assume the clinician prefers to click on the boundary point that exhibit large errors which can be quantified by Hausdorff distance (HD). We used three different head-and-neck (HN) datasets to train and test our model. For quantification, we measure the percentage of acceptable contours that meet the criterion HD95 < 2.5mm and the number of clicks required to reach this criterion.

Results: On average, 42.6% of the auto-generated contours were acceptable. With the aid of our AICIS system, the percentage of acceptable contours were improved to 92.3%. The mean number of clicks required to reach our criterion was 1.57.

Conclusion: We developed an artificial intelligence and clinician integrated systems that can substantially improve the clinician’s contour revision efficiency.

Handouts

    Keywords

    Segmentation

    Taxonomy

    IM/TH- Image Segmentation Techniques: General (Most aspects)

    Contact Email

    Share: