Author: Aditya P. Apte, Joseph O. Deasy, Jue Jiang, Nancy Lee, Sudharsan Madhavan, Nishant Nadkarni, Lopamudra Nayak, Harini Veeraraghavan, Wei Zhao 👨🔬
Affiliation: Department of Medical Physics, Memorial Sloan Kettering Cancer Center, Memorial Sloan Kettering Cancer Center 🌍
Purpose: To track early response to radiotherapy using digital twins, it is crucial to quantify tumor volume and mass changes. Traditional tumor detection methods, particularly in image registration, are often slow and imprecise. This study compares two deformable image registration (DIR) methods: the open-source conventional ANTs method and an AI-based DIR model utilizing deep learning to improve accuracy in aligning longitudinal CT images of head and neck cancer patients.
Methods: Data from 64 head and neck cancer patients were used, with two 18F-fluoromisonidazole (FMISO) CT-PET scans per patient. The first scan was taken before radiotherapy and the second, two weeks later. Images were cropped to focus on the body, and the first FMISO scan was rigidly registered to the second. Deformable registration was performed, comparing ANTs with AI-DIR model. The model was developed using a 3D convolutional long short-term memory network integrated into a U-Net encoder for abdominal organs. For head and neck, it was fine-tuned and retrained with 42 patients, including segmentations of Organs at Risk (OARs). Accuracy was evaluated by comparing deformed contours to ground truth from 22 patients using Dice similarity coefficient (DSC) and Hausdorff distance (HD95). The AI-DIR model was applied to pre-treatment CT and first FMISO scans to obtain deformed gross tumor volumes, further aligned with the second FMISO scan.
Results: AI-DIR showed slight improvements over ANTs in segmenting critical organs, achieving a Dice coefficient of 0.81 and an HD95 of 1.00 mm for the brainstem, compared to ANTs of 0.79 and 1.21 mm. In the parotid glands, AI-DIR Dice scores were 0.89 (left) and 0.88 (right), slightly better than ANTs 0.88 and 0.88.
Conclusion: Our model provides a useful and fast method to deform the tumor volumes. Further, these results will help us in calculating volume and mass loss for head and neck tumors.