Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
Presenter(s)
William Green, MD - University of Pennsylvania Perelman School of Medicine, Radnor, Pa
W. R. Green1, R. McBeth1, E. Berlin1, Z. Alexandra2, I. Alexandru3, S. Romdhani4, M. C. du Hamel de Milly4, M. Costea5, G. Temiz4, N. Paragios4, A. Chakrabarti4, and E. Ozyar6; 1Department of Radiation Oncology, University of Pennsylvania, Philadelphia, PA, 2Med Europa Hospital, Brasov, Romania, 3Institutul Oncologic “Prof. Dr. Ion Chiricu?a”, Cluj-Napoca, Romania, 4TheraPanacea, Paris, France, 5TheraPanacea, Lyon, France, 6Department of Radiation Oncology, Acibadem MAA University School of Medicine, Istanbul, Turkey
Purpose/Objective(s):
Accurate delineation of the clinical target volume (CTV) is a critical step in radiotherapy planning. In rectal cancer (RC), multiple guidelines propose different subvolumes and anatomical boundaries for pelvic lymph nodes (LN), which may lead to variability and misunderstanding in CTV definition. Artificial intelligence (AI) has the potential to standardize contouring practices. The aim of this study was to train and evaluate an AI-based model for pelvic LN segmentation according to international consensus guidelines [1].Materials/Methods:
The AI model was trained using CT images from 293 patients (134 females, 159 males), with manual delineation of 13 pelvic LN structures performed by two expert radiation oncologists according to the Valentini et al. 2016 guideline. An independent cohort of 29 patients was used for performance evaluation. An exploratory interobserver variability study was conducted on 2 patients, with both experts independently delineating all LN structures. Quantitative evaluation was performed on 31 samples per structure by comparing AI-generated and manual contours using DICE similarity coefficient and 95th percentile Hausdorff distance (HD95). Clinical acceptability was assessed on a subset of 26 patients by three expert radiation oncologists using a three-point clinical acceptability scale: A (acceptable without modifications), B (acceptable after minor corrections), and C (not acceptable for clinical use).Results:
The mean DICE across all structures was 0.72 ± 0.08, ranging from 0.57±01 for presacral abdominal LN to 0.82 ± 0.05 for right inguinal LN. Gender-based analysis demonstrated similar segmentation performance in male and female patients, with a mean DICE of 0.72 ± 0.09 and 0.72 ± 0.07, respectively. Interobserver agreement between the two experts was high for external iliac, internal iliac, and inguinal LN (DICE > 0.7), and lowest for presacral LNs (DICE < 0.3). Lower AI performance in presacral regions paralleled high interobserver variability, potentially due to anatomical ambiguity rather than model deficiency. In the qualitative evaluation, overall clinical acceptability (A+B scores) averaged 98% across all structures, ranging from 88% for presacral abdominal LN to 100% for 7 of the 13 structures. Calculated HD95 values were within clinically acceptable limits across structures.Conclusion:
A commercially available AI contouring tool trained according to international consensus guidelines demonstrated high clinical acceptability of pelvic LN contours for radiotherapy planning. This tool may improve efficiency and consistency of target delineation and may support more precise radiotherapy for pelvic malignancies. References: [1] Valentini V, Gambacorta MA, Barbaro B, et al. International consensus guidelines on Clinical Target Volume delineation in rectal cancer. Radiother Oncol. 2016;120(2):195-201. doi:10.1016/j.radonc.2016.07.017