Click here to


Are you sure ?

Yes, do it No, cancel

The Use of Artificial Intelligence to Auto-Segment Organs-At-Risk in Total Marrow Irradiation Treatment

A Liu*, R Li, C Han, J Liang, D Du, A Shinde, S Dandapani, A Amini, S Glaser, J Wong, City of Hope Medical Center, Duarte, CA


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: planning for total marrow irradiation(TMI) is a time-intensive process requiring the contouring of many organs-at-risk(OARs) throughout the entire body. This study evaluated the quality of contours auto-generated by a deep learning(DL) contouring algorithm for OARs in TMI.

Methods: first ten patients in a phase II TMI trial treated were selected for evaluation. Dose prescriptions were 20Gy to bone/lymph_nodes/spleen and 12Gy to liver/brain delivered over 5 days, twice daily. Each patient had more than 150 slices/30 structures and took approximately 6-8 hours of dosimetrist time per patient. In this study, the clinically used contours drawn by dosimetrists were used as the reference. A deep machine learning model(Ua-Net, DeepVoxel Inc) was used to auto-segment the OARs. We evaluated the performance of this DL model using 3 spatial overlap based metrics(Dice coefficient, Jaccard index(JAC) and True positive rate sensitivity(TPR)), 2 surface distance metrics(95% Hausdorff distance(HD) and average distance(AD)), 1 volume similarity index(VS). Eighteen common OARs were evaluated.

Results: DL auto-segmentation model was most similar to human generated contours for eyes, parotids, heart, liver and lungs where average Dice, JAC, TPR, HD, AD, VS were 0.85(range 0.76-0.95), 0.72(0.62-0.91), 0.85(0.74-0.98), 14.4(7.5-24.6), 4.4mm(2.0-7.9) and 0.92(0.88-0.97) respectively. Other OARs still needed improvement. Several factors contributed to the difference. The training CT used for the kidneys and spleen had patients in arms-up position, but TMI patients were simulated with arms-on-the-side. The model was trained to draw the spinal cord in contrast to the reference where spinal canal was drawn. The reference esophagus was drawn generously to minimize risk of esophagitis, whereas the model defined esophagus by CT Hounsfield boundaries.

Conclusion: auto-generated contours from a convolutional neural network model showed promise to replace human generated ones for many OARs for TMI planning, with potential to be adopted in routine clinical practice and significantly reduce the lengthy contouring time.


Treatment Planning, Bone-marrow Transplantation, Anatomical Models


IM/TH- image Segmentation: CT

Contact Email