Clinical Efficiency Gains with Artificial-Intelligence Auto-Segmentation in the Entire Human Body

Abstract

<h3>Purpose/Objective(s)</h3> To evaluate the accuracy and efficiency of Artificial-Intelligence (AI) segmentation in Total Marrow Irradiation (TMI) including contours throughout the head and neck (H&N), thorax, abdomen, and pelvis. <h3>Materials/Methods</h3> An FDA approved Artificial Intelligence (AI) segmentation software (Medical Mind, Inc) was clinically introduced for total body contouring in TMI. Our current workflow requires dosimetrist and physician to review and edit AI-segmentations prior to treatment planning. This work compares the original AI-segmentation with the clinically used segmentations including the spatial and dosimetric deviations, and the clinical efficiency gains. The TMI-model creates 27 organs at risk (OARs) and 4 planning target volumes (PTVs). Since introducing the TMI-model, 21 patients were treated with this workflow. The spatial deviations were estimated with conventional volume and surface measures including the Sorensen DICE coefficient and the 95^th% Hausdorff Distance (HD95). The dosimetric deviations were estimated for the entire dose volume histogram (DVH) for all structures using a threshold of 5% difference at any dose level. The clinical efficiency gains were estimated by the percentage of the AI-contour-surface within 1mm of the clinical contour surface. Examples of efficiency gains are: an unedited AI-contour has an efficiency gain=100%, an AI-contour with 70% of its surface <1mm from a clinical contour has an efficiency gain = 70%. <h3>Results</h3> A total of 464 contours were evaluated. In PTVs, the percentage of contour surface deviated by >1mm in 44.3%±31.1% of all targets, an average efficiency gain of 55.6%. Deviations >5mm were detected in 20.1%±34.5% of the PTV contour surfaces. In OARs, deviations >1mm were detected in 24.2%±26.9% of the structure surfaces; an average clinical efficiency gain of 75.8%. Deviations > 5mm were detected in 7.1%±17.9% of all OAR contours. DVH differences were minimal, with 17.5% of all curves exceeding 5% deviation at any dose level. The most common 5%-DVH variations were in esophagus (67%), mandible (42.9%), kidney (23.8%), and PTVs (20.2%). In H&N OARs, efficiency gains ranged from 42.0% (optic chiasm) to 100% (eyes were unedited in all cases), with mean and standard deviation 77.3%±18.9%. In thorax OARs efficiency gains were >80% in spinal cord, heart, and both lungs, with esophagus efficiency around 40% indicating significant edits. In abdomen and pelvis OARs, efficiency gains ranged from 61.6-68.6% (stomach, rectum, bowel) and 78.3-84.8% in liver and bladder. The efficiency metric was not well correlated to DICE (r = 0.75) or negatively correlated to HD95 (r=-0.75) indicating the efficiency of AI contouring is not measured by traditional overlap methods. <h3>Conclusion</h3> The Medical Mind AI software is a powerful solution for efficiency in clinical radiation therapy with the ability to segment the entire body. Based on spatial and dosimetric comparisons, structures including esophagus and PTVs require significant human edits.