ESTRO 2024 - Abstract Book

S3148

Physics - Autosegmentation

ESTRO 2024

Clinical Evaluation of a Deep Learning Model for Assessing OAR Delineation in Head and Neck Oncology

Remus C Stoica 1,2 , Adrian M Radu 3,2 , Beatrice Anghel 4 , Andrei M Dicianu 5 , Sandu Daniela-Lidia 6 , Marius Stãnescu 7 , Lucian Bicsi 7 , Dragos Dușe 7 , Dragoș Grama 7 1 Sanador, Radiation Oncology, Bucharest, Romania. 2 Synaptiq, Medical Departament, Cluj-Napoca, Romania. 3 Bucharest Institute of Oncology Prof. Dr. Alexandru Trestioreanu, Radiation Oncology, Bucharest, Romania. 4 Amethyst Radiotherapy Center, Radiation Oncology, Otopeni, Ilfov, Romania. 5 St. Nectarie Oncology and Radiotherapy Center, Radiation Oncology, Craiova, Romania. 6 Oncohelp Association, Radiation Oncology, Timisoara, Romania. 7 Synaptiq, Research Departament, Cluj-Napoca, Romania

Purpose/Objective:

Organs-at-risk (OARs) delineation, particularly in the head and neck (H&N) region, is prone to observer variability, often surpassing planning and setup errors, and is a major challenge in radiation treatment planning. Recently, deep neural networks have become increasingly popular for automated OAR contouring in radiotherapy. Metrics like volumetric or surface DICE coefficient (vDSC or sDSC) and Hausdorff distance (HD95) are now common in Artificial Intelligence (AI) discussions. However, translating these quantitative metrics into meaningful insights for radiation oncologists (ROs) is still a significant issue. This clinical evaluation aims to correlate quantitative metrics (e.g., vDSC) with qualitative assessments (clinical acceptability grades) from radiation oncologists 1 .

Material/Methods:

To compare the accuracy of the contours generated by ROs with the contours generated by AI, treatment plans from 30 patients with H&N cancer were selected to encompass approximately 25 manually delineated guidelines compliant OARs 2 . The same CT scans were used to generate contours, with the help of deep learning algorithms, for the corresponding OARs. Subsequently, the obtained 60 scans (30 with manual and the same 30 with AI-based contours) were randomly distributed to 3 different ROs to grade the H&N structures with grades ranging from 1 to 3 ( 1 – acceptable as it is, 2 – acceptable after minor corrections, 3 – acceptable after major corrections ) 3 . The grading criteria adhered to the peer review guidelines established by the Royal College of Radiologists (RCR). The qualitative analyses - represented by these grades - were then subjected to correlation with the quantitative analysis obtained through AI-generated contours.

Results:

In the qualitative analysis conducted by the 3 independent ROs, AI-based contouring outperformed manual methods for 14 out of 25 OARs, judging by the proportion of contours graded with 1. For two particular structures (Brain and Eye_R), the performance of AI and manual contouring was equivalent. In terms of quantitative scores, the median vDSC surpassed 0.80 for 19 out of the 24 structures analyzed, with the "Body" structure being excluded from this assessment. Notably, the "Brain," "Bone_Mandible" and "Eye_L/R" structures demonstrated outstanding median vDSC values of 0.99, 0.93, and 0.93, respectively. On the other hand, "OpticChiasm" presented challenges for both AI and RO due to poor visualization in the CT images, resulting in a lower median vDSC of 0.65. Despite this, the achieved score is considered favorable when compared to existing literature, which reports scores ranging from 0.37 to 0.63 for this particular structure 4,5,6 .