ESTRO 2024 - Abstract Book

S3102

Physics - Autosegmentation

ESTRO 2024

Pathologically validated deep learning model for laryngeal and hypopharyngeal GTV delineation on MRI

Koen M. Kuijer 1,2 , Hilde J.G. Smits 1 , Patricia A.H. Doornaert 1 , Kenan Niu 3 , Ernst J. Smid 1 , Chris H.J. Terhaard 1 , Mischa de Ridder 1 , Marielle E.P. Philippens 1 1 University Medical Center Utrecht, Department of Radiotherapy, Utrecht, Netherlands. 2 University of Twente, Technical Medicine, Enschede, Netherlands. 3 University of Twente, Robotics and Mechatronics group, Faculty of Electrical Engineering, Mathematics and Computer Science, Enschede, Netherlands

Purpose/Objective:

Accurate and reproducible target delineation is the cornerstone of modern radiotherapy. Deep learning is a promising approach to increase reproducibility and time-efficiency of GTV delineation in head and neck cancer. Currently, most models use CT and FDG-PET image input, despite the superior soft tissue contrast of MRI. Additionally, current model evaluation primarily relies on manually delineated GTVs as reference, which are subjective and tend to overestimate tumor volume[1]. In contrast, tumor delineations on histopathology exhibit a high level of consistency[2] and are considered the gold standard for tumor delineation. Therefore, the aim of this study was to develop an MRI-only deep learning model for laryngeal and hypopharyngeal GTV delineations and validate the model with pathological tumor delineation.

Material/Methods:

A retrospective clinical dataset of 193 patients with laryngeal or hypopharyngeal carcinoma treated with radiotherapy was divided into a training (n=155) and test (n=38) set. We trained an nnUNet[3] model with three fat-suppressed MRI sequences (T1-weighted gadolinium enhanced, T2-weighted and DWI) as input, and the clinically used GTV delineations as ground truth. Two separate datasets were used to evaluate the model’s performance: the clinical test set and a pathology data set. The latter contains 18 patients who underwent a total laryngectomy. Manual GTV delineations were made on the MRI acquired before surgery. For 16 patients in the laryngectomy group, the surgical specimen was histologically processed, virtually reconstructed in 3D and registered to the in vivo MRI, allowing for a pathological tumor delineation that serves as a ground truth in this dataset[4]. In both datasets, quantitative assessment between automatic and manual GTV delineations was done using the Dice similarity coefficient (DSC), and type I (underestimation) and type II (overestimation) 95th-percentile Hausdorff distance (HD95). For pathological validation, the automatic and manual GTV delineations were compared to the pathological tumor delineation using the sensitivity, positive predictive value (PPV), volume overestimation, and type I HD95. The latter is used as a measure for the required CTV margin to obtain adequate tumor coverage. For qualitative evaluation, four radiation oncologists scored the automatic and manual delineations in the pathology dataset for clinical acceptability for treatment planning without the need for adjustments.

Results:

In the clinical test set, the median DSC and type I and II HD95 values between the automatic and manual delineations were 0.58, 6.0 mm and 3.0 mm respectively. In the pathology dataset, the model performance substantially improved, with a median DSC, type I HD95, and type II HD95 of 0.81, 2.0 mm, and 2.9 mm respectively. This improvement can be largely attributed to the absence of registration errors due to the use of MRI as the reference image for manual delineations in the pathology dataset, as opposed to CT in the clinical dataset.