MedMLX: On-Device Medical Imaging AI for Apple Silicon
Abstract
Background: Medical imaging research increasingly combines specialized artificial intelligence models for segmentation, detection, image synthesis and report generation. For three-dimensional CT and MRI, these workflows typically run on NVIDIA GPUs with substantial memory. On-device AI on widely available Apple Silicon computers could broaden access to these models for clinical researchers. These computers share a large memory pool between the CPU and GPU, which suits workflows that combine several models. However, few established imaging models run natively on Apple Silicon. Adapting a model to new hardware also requires verifying that its results match the original software. Methods: We developed MedMLX, a collection of 21 established medical imaging models rewritten for Apple Silicon in MLX, Apple's machine learning framework. The collection includes nnU-Net, TotalSegmentator, MedSAM2 and SAM 3. The models segment organs and tumors, detect lung nodules, generate synthetic images and draft radiology reports from CT, MRI, PET/CT, radiographs and endoscopic images. Using public cases from 11 imaging datasets, we compared the models' results with those of the original software run on an Intel CPU and an NVIDIA GPU. Agreement required a Dice coefficient of at least 0.99 for every segmentation label and equivalent criteria for other outputs. We measured memory use and processing time on an Apple M1 Max computer with 64 GB of memory. Results: MedMLX agreed with the original software in 211 of 212 comparisons. The single exception was a brain MRI segmentation. On this case, the original software itself gave different results on the two devices. We illustrated sequential workflows, including anatomical segmentation of a synthetic brain MRI and lesion synthesis guided by organ segmentation. With a memory-saving setting, 19 of the 21 models used less than 16 GB of memory. This setting reduced memory use by a median of 49% and increased processing time by a median of 13%. Conclusions: MedMLX enables on-device AI with established medical imaging models on Apple Silicon computers rather than specialized NVIDIA hardware. The models agree closely with the original software, and most need less than 16 GB of memory.
Related
- Early applications of apple vision pro in clinical medicine and medical education: a scoping review
- Multi-Contrast Deep-Silicon Photon-Counting CT for Post-Embolization Imaging: A Quantitative Phantom Study
- Virtual Noncontrast Imaging with Deep-Silicon Photon-Counting CT: Matched Phantom and Initial Human Evaluation
- Does population aging promote embeddedness in the regional medical device trade network? Evidence from RCEP member countries
- CE Mark Is Not Enough: Medical Device Safety in African Healthcare Systems
BibTeX
@misc{sandoval2026medmlx,
title = {{MedMLX: On-Device Medical Imaging AI for Apple Silicon}},
author = {A. Sandoval and J. Sandoval and A. Saleem},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.28.26366762},
url = {https://www.medrxiv.org/content/10.64898/2026.10.28.26366762v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - MedMLX: On-Device Medical Imaging AI for Apple Silicon
AU - A. Sandoval
AU - J. Sandoval
AU - A. Saleem
PY - 2026
DA - 2026/10/28
PB - medRxiv
DO - 10.64898/2026.10.28.26366762
UR - https://www.medrxiv.org/content/10.64898/2026.10.28.26366762v1
AB - Background: Medical imaging research increasingly combines specialized artificial intelligence models for segmentation, detection, image synthesis and report generation. For three-dimensional CT and MRI, these workflows typically run on NVIDIA GPUs with substantial memory. On-device AI on widely available Apple Silicon computers could broaden access to these models for clinical researchers. These computers share a large memory pool between the CPU and GPU, which suits workflows that combine several models. However, few established imaging models run natively on Apple Silicon. Adapting a model to new hardware also requires verifying that its results match the original software. Methods: We developed MedMLX, a collection of 21 established medical imaging models rewritten for Apple Silicon in MLX, Apple's machine learning framework. The collection includes nnU-Net, TotalSegmentator, MedSAM2 and SAM 3. The models segment organs and tumors, detect lung nodules, generate synthetic images and draft radiology reports from CT, MRI, PET/CT, radiographs and endoscopic images. Using public cases from 11 imaging datasets, we compared the models' results with those of the original software run on an Intel CPU and an NVIDIA GPU. Agreement required a Dice coefficient of at least 0.99 for every segmentation label and equivalent criteria for other outputs. We measured memory use and processing time on an Apple M1 Max computer with 64 GB of memory. Results: MedMLX agreed with the original software in 211 of 212 comparisons. The single exception was a brain MRI segmentation. On this case, the original software itself gave different results on the two devices. We illustrated sequential workflows, including anatomical segmentation of a synthetic brain MRI and lesion synthesis guided by organ segmentation. With a memory-saving setting, 19 of the 21 models used less than 16 GB of memory. This setting reduced memory use by a median of 49% and increased processing time by a median of 13%. Conclusions: MedMLX enables on-device AI with established medical imaging models on Apple Silicon computers rather than specialized NVIDIA hardware. The models agree closely with the original software, and most need less than 16 GB of memory.
ER -