Abstract
Multimodal CXR-laboratory models are usually evaluated on server GPUs, but local deployment may be needed where connectivity, data governance requirements, or hardware budgets limit cloud inference. We tested whether CXR-plus-laboratory mortality models trained centrally could be executed locally on low-cost edge hardware. Using the Symile-MIMIC dataset, we trained ResNet-18 and MobileNetV3-Small late-fusion models for early-hospitalization mortality prediction and compared inference performance across A100 GPU (PyTorch), Raspberry Pi 5 CPU (TorchScript), and Raspberry Pi AI HAT+ 2 (NPU) acceleration. ResNet-18 preserved discrimination across platforms, achieving AUROC values of 0.8005 on the A100 GPU and 0.7989 on the AI HAT+ NPU. The accelerator reduced the Raspberry Pi’s CPU mean latency from 423.21 ms to 9.98 ms on NPU and reduced the model artifact size from 43.37 MiB to 8.19 MiB. MobileNetV3-Small reduced CPU latency but failed to complete accelerator compilation. These results support the technical feasibility of local multimodal model-call inference on this Raspberry Pi-based hardware stack.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_099.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=w55P5QVgW3
BibTex
@InProceedings{VilIsm_Edge_MICCAISAT2026,
author = { Villanueva-Miranda, Ismael AND Xiao, David AND Gu, Zifan},
title = { { Edge Deployment of Multimodal Clinical AI on Low-Cost Hardware for Resource-Constrained Settings } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}