3115 - Using Artificial Intelligence to Plan Total Marrow/Lymphoid Irradiation (TMI/TMLI): A Feasibility Study

Purpose/Objective(s): Total marrow irradiation (TMI) and total marrow & lymphoid irradiation (TMLI) are being considered as targeted forms of total body irradiation (TBI) conditioning for patients undergoing hematopoietic stem cell transplantation. While TMI/TMLI have advantages over TBI, such as improved disease control and reduced risk of secondary cancers, the longer planning time of TMI/TMLI is a major drawback. While most TMI/TMLI targets, such as long bones and the liver, are relatively simple to contour, several targets, like ribs and vertebrae, are technically complex. Preliminarily, contouring the ribs and vertebrae of one patient takes us at least one hour and at least 110 minutes, respectively. Moreover, lymph nodes—targets in TMLI—have significant variability in size/location. We evaluated the feasibility of using artificial intelligence (AI)-powered, “auto” contouring software to decrease the time spent on generating a TMI/TMLI plan without compromising accuracy. We hypothesized the use of automatic contouring, even with manual correction, would require significantly less time than that required for manual contouring the same plan, thereby making this conditioning strategy more clinically feasible.

Materials/Methods: Using de-identified chest CT scans (n = 3), we compared two contouring workflows: the mean total time required for a radiation oncologist to manually contour lymph nodes, ribs, vertebrae, lungs, liver, spleen, submandibular glands, and thyroid versus the time required for medical students (untrained operators) to use automatic contouring software to contour the same structures. Time required for a radiation oncologist to manually correct the AI’s contouring was included. All contours (n = 6) were then reviewed for accuracy by a radiation oncologist. Total time spent on contouring was compared across workflows using the unpaired t-test (a = 0.05; GraphPad).

Results: Three untrained operators used AI software to each contour 1 unique chest CT scan. Times required for automatic contouring were 00:08:35, 00:12:13, and 00:05:57 (mean = 00:08:55 ± 00:03:09); times required for manual correction were 01:07:55, 01:11:30, and 01:06:42, respectively (mean = 01:08:42 ± 00:02:30). Mean total time required for automatic TMI/TMLI contouring was 01:17:37 (± 00:05:37). A radiation oncologist manually contoured the same CT scans and the total times for contouring were 03:59:28, 04:25:39, and 04:28:01 (mean = 04:17:43 ± 00:12:56). Unpaired t-test revealed automatic contouring was significantly faster (p < 0.0001) than manual contouring.

Conclusion: AI-powered automatic contouring software can contour CT scans for a TMI/TMLI plan with over a 3-fold reduction in total time, without loss of accuracy, compared to manual contouring by a radiation oncologist. Thus, automatic contouring software is a practical tool that could make TMI/TMLI more clinically feasible.