2867 - Machine-Learning Algorithm for Head and Neck Cancer Nodal Volume Segmentation: Expert vs. Learner Comparison
Presenter(s)
J. Y. Wang1, D. Provenzano2, S. Goyal1, and Y. J. Rao1; 1Department of Radiation Oncology, George Washington University School of Medicine and Health Sciences, Washington, DC, 2Biomedical Engineering, George Washington University School of Engineering and Applied Science, Washington, DC
Purpose/Objective(s): Machine learning algorithms automate the contouring of nodal volumes for head and neck cancer radiotherapy planning. This study compares the performance of an expert clinician, clinician learner, and auto-contouring algorithm in lymph node target delineation.
Materials/Methods: Data was retrospectively collected from 10 consecutive patients treated with definitive radiotherapy for head and neck cancer between 2024 and 2025 (IRB number NCR191470). Clinical target volumes (CTVs) for intensity-modulated radiotherapy for lymph node anatomical regions IB, II, III, IV, and V (left and right as separate structures) were generated by (1) manual contouring by an expert (attending physician with >5 years’ experience), (2) clinician learner (medical student), and (3) algorithms trained separately by both the expert and learner. The algorithm was a multi-atlas auto-contouring algorithm implemented in a commercial treatment planning system. Algorithm training was bootstrapped by training on patients 1-3 and then using model generated contours with manual editing for each subsequent patient. These were then placed back into the training set to create an updated model that was iteratively improved upon all 10 patient scans. Three sets of trials were analyzed: (1) manual contouring (trials 1-3), (2) manually editing machine-learning contours (trials 4-7), and (3) minor modifications to the algorithm’s contours (trials 8-10). Manual contouring and editing times were measured for each patient. Correlation index and dice coefficient between expert and learner contours were measured and compared (mean±SD).
Results: Contours generated by the machine learning algorithms universally saved time. The average time required for manual contouring for the expert and learner were 14.2 and 34.6 minutes, respectively. Once trained with the information from the 3 manual contours, the ML algorithm required significantly less time for preparing treatment plans during trials 4-7 (6.8 and 34.6 minutes for the expert and learner, respectively) and even less time by trials 8-10 (3.3 and 9.4 minutes for the expert and learner, respectively). In terms of accuracy, the ML algorithm performed similarly compared to the expert when utilizing the expert’s contours and separately, when utilizing the learner’s contours. Correlation coefficient between expert and learner contours improved from 1.14±0.10 with manual contouring (trials 1-3) to 1.11±1.11 for the fully trained model (trials 8-10) (p=0.04). Dice coefficient improved from 0.88±0.08 to 0.89±0.08, but did not reach statistical significance (p=0.06).
Conclusion: The ML algorithm demonstrated strong potential for improving efficiency in head and neck cancer treatment planning, reducing contouring time for lymph nodes IB through V significantly when compared to manual methods. The algorithm allowed the expert to produce nodal target volumes within about 3 minutes and is now in clinical use at our institution.