MOZI TPS

The MOZI TPS is an advanced software system designed to facilitate external beam photon radiotherapy planning by integrating image processing, deep learning-driven automatic structure delineation, and Monte Carlo dose calculation to generate optimized treatment plans. Its methodology centers on providing clinicians with a comprehensive suite of automated and manual tools, with its performance in image registration, contouring accuracy, and dose computation for IMRT/VMAT validated through rigorous testing against established systems and standards.

The MOZI Treatment Planning System (MOZI TPS) is a standalone software application designed for planning radiotherapy treatments for patients with both malignant and benign diseases, specifically for external beam irradiation using photon beams. Its core functionalities encompass a comprehensive workflow including image processing with automatic rigid and deformable registration, structure delineation utilizing deep learning for automatic contouring of organs at risk, treatment plan design, dose calculation via the Monte Carlo method, plan optimization for techniques like IMRT and VMAT, and subsequent plan evaluation. The system also provides user login, a graphical user interface, and tools for system and patient data management, supporting the DICOM RT standard for interoperability with other radiotherapy equipment and systems. The MOZI TPS employs established algorithms for image registration and incorporates a Monte Carlo based algorithm for photon dose calculation (1MeV-20MeV range), which has been validated for accuracy. Its automatic segmentation capabilities are driven by deep learning models, trained on Chinese datasets and tested on independent US datasets covering various anatomical sites (Head & Neck, Thorax, Abdomen, Pelvis) and demonstrating performance with Dice Similarity Coefficients provided for numerous OARs. Performance testing confirmed its dose calculation, image registration (using NMI and HD metrics), and structure delineation capabilities against predicate and reference devices. Comprehensive end-to-end testing on 18 patient cases, including lung, liver, brain, and Head & Neck cancers using IMRT and VMAT techniques, validated the entire clinical workflow from CT image import to plan export and phantom measurement validation, ensuring the system meets safety (IEC 62083-2009) and performance standards (AAPM TG-119).

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. The document mentions deep learning for automatic contouring and provides Dice Similarity Coefficient (DSC) values for various OARs. What specific deep learning architecture is employed, and how does its performance on particularly challenging cases (e.g., post-operative anatomy or re-irradiation scenarios with altered tissue characteristics) compare to manual contouring by experienced clinicians in terms of both accuracy and time efficiency?; The Monte Carlo dose calculation algorithm is stated to be verified. Could you provide more details on the specific beam modeling parameters, variance reduction techniques utilized, and how its calculation speed for complex IMRT/VMAT plans compares to the predicate's AAA/AXB algorithms without compromising dosimetric accuracy, especially for heterogeneous tissues?; The end-to-end testing involved 18 patient cases for lung, liver, brain, and Head & Neck cancers. Can you elaborate on the diversity of these cases beyond the listed cancer types, specifically concerning tumor stage, patient anatomical variations, the inclusion of motion management strategies, and the complexity of planning objectives (e.g., simultaneous integrated boost, or targets near highly sensitive organs)?; For the deformable image registration, Normalized Mutual Information (NMI) and Hausdorff Distance (HD) were used for evaluation. Beyond these quantitative metrics, what processes or qualitative assessments are in place to validate the clinical plausibility and anatomical correctness of the deformations, especially in regions with significant pathological changes or low image contrast?; The training data for CT structure models was sourced from China, while the testing data was from the United States, with specified demographic and scanner diversity. What specific strategies were implemented during model development and validation to ensure robust generalization across different populations, varied imaging protocols, and potential biases introduced by scanner vendor differences, and were any domain adaptation techniques employed?