Paper Proffered Program Therapy Physics

Contourless Patient-Specific AI Triggering of Online Adaptive IMPT In Oropharyngeal Cancer

Abstract

Purpose

Online adaptive proton therapy is limited by the time required for target/OAR re-segmentation and dosimetric review. We developed a contourless patient-specific AI framework to automatically trigger online plan adaptation in oropharyngeal cancer (OPC) patients treated with IMPT.

Methods

An in-house Dual-CNN model predicting optimal OPC treatment isocenter position served as the base model. Data from 40 OPC-IMPT patients were analyzed, each with one planning CT (pCT) and ~35 daily CBCT-based synthetic CTs (sCTs) per patient. 14/40 patients required clinical plan adaptation. MOQUI was employed for dose calculations. For each patient, the pCT was augmented using 80 random isocenter shifts (±15mm) and corresponding dose distributions used to transfer-learn the base model into a patient-specific model before treatment. On each fraction, the sCT was augmented using 16 random shifts (±7mm), and resulting dose distributions were used to predicted isocenter positions. The standard deviation (SD) of predicted-isocenter positions on sCT (σsCT) was compared to the pCT baseline (σpCT); the relative difference (Δσ) served as the trigger signal (Δσ>60%). At patient-level, maximum Δσ per patient was used to discriminate adaption need. At fraction-level, a 5-fraction prediction horizon was predefined: fractions were labeled positive if clinical adaptation occurred within the subsequent five fractions and negative otherwise. Sensitivity, specificity, ROC/AUC, and workflow runtime were evaluated.

Results

13/14 adaptive patients were correctly identified (sensitivity=0.93). 5/26 non-adaptive patients were flagged as adaptive (specificity=0.81), predominantly reflecting transient anatomical changes with clinically acceptable coverage. Δσ was higher in positive versus negative fractions (133.2% vs 24.1%). ROC/AUC demonstrated strong discrimination at patient/fraction levels (0.94/0.92, respectively). End-to-end processing completed in <2 min per fraction.

Conclusion

A contourless patient-specific AI trigger can identify patients and fractions likely to benefit from online IMPT adaptation by monitoring changes in treatment-isocenter prediction standard deviation relative to baseline, without target re-segmentation or DVH review.

People

Hooman BahrdoPresenting Author · Department of Radiation Oncology, University Medical Center Groningen, University of Groningen, Groningen, The Netherlands Gabriel Gutteres Marmitt, PhDAuthors · Department of Radiation Oncology, University Medical Center Groningen, University of Groningen, Groningen, The Netherlands Femke Oosterhof, PhDAuthors · Department of Radiation Oncology, University Medical Center Groningen, University of Groningen, Groningen, The Netherlands Pietro Pisciotta, PhDAuthors · Department of Radiation Oncology, University Medical Center Groningen, University of Groningen, Groningen, The Netherlands Makbule Atasoyu, PhD, MDAuthors · Department of Radiation Oncology, University Medical Center Groningen, University of Groningen, Groningen, The Netherlands Daniele Cannavo, PhDAuthors · Department of Radiation Oncology, University Medical Center Groningen, University of Groningen, Groningen, The Netherlands Johannes A. Langendijk, PhD, MDAuthors · Department of Radiation Oncology, University Medical Center Groningen, University of Groningen, Groningen, The Netherlands Stefan Both, PhDAuthors · Department of Radiation Oncology, University Medical Center Groningen, University of Groningen, Groningen, The Netherlands