Click here to

Session: Data Science Autoplanning and Autosegmentation [Return to Session]

A Sensitivity Analysis On the Relationship Between Dose and Overlap Metrics for Head & Neck Normal Tissues

B Marquez1,2*, C Owens1,2, K Huang1,2, M El Basha1,2, R Mumme2, C Nguyen2, C Peterson1,2, D Fuentes1,2, T Whitaker1,2, T Netherton1,2, L Court1,2,(1) University of Texas MD Anderson Cancer Center UTHealth Graduate School of Biomedical Sciences, Houston, TX, (2) MD Anderson Cancer Center, Houston, TX


SU-H430-IePD-F5-3 (Sunday, 7/10/2022) 4:30 PM - 5:00 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 5

Purpose: To understand the clinical relevance (i.e., impact on tissue dose) of geometric indices like dice similarity coefficient (DSC).

Methods: A clinically deployed, auto-contouring model is routinely used in our institution’s head & neck (H&N) service and has auto-contoured 1562 patients to date (up to 21 normal tissues per patient). H&N normal tissue auto-contours are used as-is or edited prior to final plan approval. In this study, we used DSC, surface DSC (2mm tolerance) and Hausdorff distance (HD) to quantify the extent of physician edits over 31 retrospectively treated patient cases. We then evaluated the impact of the edits on reported dose by comparing the mean and maximum doses to the auto-contoured versus edited structures against their respective geometric indices.

Results: Overall, 90% of all geometric comparisons between automatic and edited structures scored a DSC>0.78, surface DSC>0.89, and HD<8mm. Dosimetrically, the difference between mean and max dose to auto- and edited-contours was less than 50cGy for 90% of structures. There was a clear relationship between DSC and surface DSC for structure edits that resulted in a dose difference >= 200cGy (~3% of the dose prescription). The percentage of structures with max dose differences larger than 200cGy was 43%, 14%, and 3% for surface DSC ranges of 0.0-0.5, 0.6-0.8 and 0.9-1, respectively, with corresponding percentage for DSC as 38% , 7%, and 2%. The percentage with mean dose differences larger than 200 cGy were 86%, 42%, and 1% for surface DSC ranges of 0.0-0.5, 0.6-0.8, and 0.9-1, respectively, with corresponding percentage for DSC as 50%, 16%, and 2%.

Conclusion: Most edits to auto-contoured H&N normal tissues are minor with minimal effect on doses calculated to those tissues. For edits that give a meaningful dose difference, we were able to identify a relationship between the calculated tissue dose and geometric metrics.

Funding Support, Disclosures, and Conflict of Interest: Funding: Cancer Prevention and Research Institute of Texas


Segmentation, Contour Extraction, Dose


IM/TH- image Segmentation: CT

Contact Email