ATOMMIC: An Advanced Toolbox for Multitask Medical Imaging Consistency to facilitate Artificial Intelligence applications from acquisition to analysis in Magnetic Resonance Imaging

TL;DR

ATOMMIC presents an open-source multitask toolbox that unifies MRI tasks across reconstruction, segmentation, and quantitative parameter mapping, addressing generalization through harmonized complex-valued and real-valued data handling and an integrated Multitask Learning framework. The authors benchmark 25 DL models on 8 public datasets, demonstrating that physics-informed models enforcing data consistency outperform purely data-driven approaches, and that end-to-end MTL can improve both reconstruction and segmentation. By providing standardized data loaders, transforms, and training pipelines, ATOMMIC enables robust cross-task benchmarking and reproducible research, with pretrained checkpoints and comprehensive documentation. The framework holds potential to accelerate MRI AI development in acquisition-to-analysis workflows, while highlighting areas for privacy-aware private-data integration and expansion to additional tasks and modalities.

Abstract

AI is revolutionizing MRI along the acquisition and processing chain. Advanced AI frameworks have been developed to apply AI in various successive tasks, such as image reconstruction, quantitative parameter map estimation, and image segmentation. Existing frameworks are often designed to perform tasks independently or are focused on specific models or datasets, limiting generalization. We introduce ATOMMIC, an open-source toolbox that streamlines AI applications for accelerated MRI reconstruction and analysis. ATOMMIC implements several tasks using DL networks and enables MultiTask Learning (MTL) to perform related tasks integrated, targeting generalization in the MRI domain. We first review the current state of AI frameworks for MRI through a comprehensive literature search and by parsing 12,479 GitHub repositories. We benchmark 25 DL models on eight publicly available datasets to present distinct applications of ATOMMIC on accelerated MRI reconstruction, image segmentation, quantitative parameter map estimation, and joint accelerated MRI reconstruction and image segmentation utilizing MTL. Our findings demonstrate that ATOMMIC is the only MTL framework with harmonized complex-valued and real-valued data support. Evaluations on single tasks show that physics-based models, which enforce data consistency by leveraging the physical properties of MRI, outperform other models in reconstructing highly accelerated acquisitions. Physics-based models that produce high reconstruction quality can accurately estimate quantitative parameter maps. When high-performing reconstruction models are combined with robust segmentation networks utilizing MTL, performance is improved in both tasks. ATOMMIC facilitates MRI reconstruction and analysis by standardizing workflows, enhancing data interoperability, integrating unique features like MTL, and effectively benchmarking DL models.

Figures (10)

• Figure 1: Overview of AI repositories for MRI tasks parsed from GitHub. The repositories are divided into two groups on the x-axis: those that do not support complex-valued data (left group) and those that do support complex-valued data (right group). The repositories are further categorized based on their activity level, on the top of the x-axis, split by vertical dashed lines: those that have committed updates within the last year (2023) are labeled as "Active Yes" (right side), and those that have not as "Active No" (left side). The supported tasks are depicted on the y-axis, while the horizontal dashed line showcases the multitasking toolboxes. Repositories are visualized as dots if documentation is available (Docs Yes) and cross marks if documentation is unavailable (Docs No). Each color signifies the language of the repository, with blue representing C++, brown representing Julia, gray representing MATLAB, and green representing Python.
• Figure 2: Schematic overview of ATOMMIC. Starting from the left to the right, MRI data are given as input. Next, configurations such as dataloaders, undersampling schemes, transforms, task(s) and models, optimizers, learning rate schedulers, losses, training and optimization settings, evaluation metrics, and exports are defined. The output is an atommic artifact (rightmost) containing the trained model's checkpoints and configurations, which can be directly used for inference on new datasets.
• Figure 3: Overview of undersampling options in ATOMMIC. From left to right, columns one to seven present retrospective undersampling using Equispaced 1D (E1D), Equispaced 2D (E2D), Gaussian 1D (G1D), Gaussian 2D (G2D), Random 1D (R1D), Poisson 2D (P2D), and Poisson 2D with 20% Partial Fourier (P2DPF) masking, respectively. Note that Partial Fourier can be applied to any masking. The last column presents prospective undersampling (Prosp) using the Calgary-Campinas 359 dataset default 2D Poisson mask (beauferrisMultiCoilMRIReconstruction2022).
• Figure 4: Multicoil-related transforms applied to example data from the CC359 dataset (beauferrisMultiCoilMRIReconstruction2022). Fig. \ref{['fig:cc359-12-coils']} shows the fully sampled 12-coil Ground Truth (GT) data. Fig. \ref{['fig:cc359-csm-all-coils']} shows the estimated 12-coil Coil Sensitivity Maps (CSM) using the EstimateCoilSensitivityMaps class. In Fig. \ref{['fig:cc359-gdcc']}, the 12-coil data are reduced to 4 coils after using the Geometric Decomposition Coil Compression (GDCC) method. Fig. \ref{['fig:cc359-gt-rss-sense-gdcc']} shows different Coil Combination Methods (CSM), such as the Root-Sum-of-Squares (RSS) and the SENsitivity Encoding (SENSE), applied to the Ground Truth (GT) (first and second, respectively), and the GDCC (third and fourth, respectively).
• Figure 5: Reconstructions of 12-coil T$_1$-weighted data from the CC359 dataset, undersampled with a Poisson disc distribution 2D sampling pattern for 5x (Fig. \ref{['subfig:recon_cc359-5x']}) and 10x (Fig. \ref{['subfig:recon_cc359-10x']}) acceleration. The top row-first column shows the ground truth (Target) image. SSIM and PSNR scores are reported for each method and computed against the Target image. Methods are sorted alphabetically.