ESTRO 2024 - Abstract Book

S1420

Clinical - Head & neck

ESTRO 2024

In patients with oropharyngeal cancer treated with radiation therapy, the choice of contralateral irradiation is often driven by the clinical proximity to midline and the extent of nodal disease. However, some patients with oropharynx cancer may have decreased contralateral irradiation by volume rather than complete omission of contralateral treatment. Retrospective methods to objectively assess the extent of contralateral irradiation in relationship to objective primary tumor and nodal features may help identify patients who would benefit or be at risk via omission of contralateral irradiation. In order to objectively assess the extent of lymph node irradiation in large cohorts of patients, automated objective methods to assess elective lymph node treatment are crucial. In this work, we demonstrate a novel approach using a convolutional neural network (CNN) to autosegment, retrospectively label, and classify the extent of elective neck irradiation in previously treated patients and analyze the relationship with clinical and patient features. Initially, 100 patients with head and neck cancer were contoured to identify all elective neck level targets as per standard guidelines 1 , including levels IA, IB, II, III, IV, V, VIIA on both left and right necks. These patients were used to train a CNN neck level model (NLM) which was validated on a further 10 patients clinically. After training, the resultant NLM was applied retrospectively to all patients in the RADCURE dataset 2 . Overlap between the autosegmented NLM model contours with existing elective neck CTV targets was assessed using the proportion of voxel overlap of the individual neck structures with the elective CTV which was clinically defined. The original clinically defined primary gross tumor volumes (GTVp) and nodal volumes (GTVn) were contoured by treating radiation oncologists. Objective tumor laterality (OTL) was calculated as the difference between the GTVp center of mass and the midline in the R/L direction. OTL accuracy compared to the clinically assigned laterality (when available) was assessed. OTL was used to normalize right and left neck targets to ipsilateral and contralateral relative to the primary tumor. Midline proximity (MP) was the difference between the medial edge of the GTVp bounding box to the patient midline with positive values representing progressively lateralized tumors and negative values representing tumors crossing midline. Neck levels were assumed to be 'treated’ if >70% of the auto segmented volume was overlapped with the clinical elective CTV target in the respective neck. Summary statistics were calculated by tumor volume, midline proximity, staging information, and clinical variables to evaluate patterns of CTV coverage and associations with tumor or patient factors. A total of 1142 patients with oropharyngeal squamous cell carcinoma were identified from within the RADCURE dataset and were included in the analysis as they had the primary tumor laterality defined clinically. After NLM application, comprehensive bilateral neck irradiation (IA-VIIA) occurred in 33% of patients. In patients with oropharyngeal primary disease, contralateral level II overlap showed approximately 12% of oropharynx patients were treated with ipsilateral intent only consistent with our previous reports from our institution. 3 In the oropharynx, multivariable LR demonstrated GTVp volume (p<0.01) and ipsilateral nodes (p<0.01) were associated with gross contralateral nodal involvement. Neither HPV status or subsite (e.g. base of tongue) was significant on multivariable analysis. Of the patients without contralateral nodal disease (n=98), the mid-line proximity and ipsilateral nodal involvement were both associated with increased use of contralateral elective nodal irradiation as assessed using the NLM overlap method. Material/Methods: Results:

Conclusion:

Automatic assessment of elective contralateral neck coverage using a neck level auto-segmentation model shows relationships between primary tumor features including midline proximity and elective target . Objective methods