Advanced Algorithms for Treatment Management Applications (AATMA)

AATMA™ is a medical image processing library that leverages machine-learning convolutional neural networks via an API to perform auto-segmentation of 3D CT images for radiation therapy planning. Its core function is to produce derived datasets for treatment planning systems, with its efficacy demonstrated through validated performance of Head & Neck and Male Pelvis models, necessitating review by qualified healthcare professionals before clinical application.

Advanced Algorithms for Treatment Management Applications (AATMA™) is a medical image processing library designed to generate derived datasets for radiation therapy treatment planning systems and other pre-treatment planning applications. It functions as an accessory, utilizing machine-learning convolutional neural networks for the auto-segmentation of medical images, and includes pre-trained models specific to body sites and imaging modalities. AATMA™ operates as a computational engine without storing data or logs, lacks a user interface, and is accessed via an Application Programming Interface (API), requiring external software for the review and editing of automatically generated contours. All datasets produced by AATMA™ must be reviewed and validated by a qualified healthcare professional prior to clinical use. Performance of AATMA™ was evaluated through non-clinical software verification and validation, specifically for two pre-trained models utilizing 3D CT image sets: Head & Neck and Male Pelvis. The Head & Neck model, trained on 66 image sets, achieved an average DICE coefficient of 0.84 during verification with 6 different patient CTs and passed validation with 13 additional image sets. The Male Pelvis model, trained on 205 image sets, demonstrated an average DICE coefficient of 0.93 in verification using 5 patient CTs and also met validation criteria with 20 further image sets, confirming its ability to meet user needs for auto-segmentation.

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. What specific anatomical structures are the Head & Neck and Male Pelvis models trained to segment, and what is the variability in performance (e.g., DICE coefficient) across these different structures within each model?; How does the performance of AATMA™, particularly segmentation accuracy and processing time, vary when applied to CT images from different scanner manufacturers or with diverse acquisition parameters not included in the training/validation sets?; What are the mechanisms or recommended best practices for clinical sites to validate AATMA™'s pre-trained models with their local patient data and imaging protocols before routine clinical use?; Given that AATMA™ is an API-driven library, what are the detailed computational requirements (e.g., GPU specifications, memory, processing cores) for integrating and optimally running the auto-segmentation algorithms within various existing treatment planning systems?; Are there plans to expand the AATMA™ library with pre-trained models for other anatomical regions, different imaging modalities (e.g., MRI, PET-CT), or to address specific pathological conditions beyond standard anatomical segmentation?