Abstract
Remote and resource-constrained settings face persistent barriers to accessing specialized prenatal imaging, including shortages of trained sonographers, limited ultrasound availability, and inadequate computing infrastructure. Although artificial intelligence has shown strong performance for fetal plane classification, most approaches are evaluated on curated datasets or require cloud-based inference and high-capacity hardware, limiting their use in point-of-care ultrasound workflows. This study benchmarks ten lightweight and heavyweight CNN- and transformer-based architectures under mobile deployment constraints. The models were pretrained on a public fetal ultrasound dataset, fine-tuned on a phantom-based blind-sweep dataset, exported to ONNX, and optimized through pruning and INT8 quantization. Performance was evaluated on low-, mid-, and high-tier Android devices using macro-averaged F1-score, p90 inference latency, and FPS throughput. On low-tier hardware, optimized ResNet34 achieved a macro-averaged F1-score of 0.88, a p90 latency of 85.2 ms, and a throughput of 15.64 FPS, indicating a favorable classification–efficiency trade-off. In comparison, Swin-T achieved a similar F1-score of 0.87 but showed substantially higher p90 latency and lower throughput, at 272.9 ms and 4.12 FPS, respectively. These results demonstrate that similar classification performance can lead to markedly different mobile deployment characteristics. Overall, optimized convolutional architectures showed more favorable trade-offs for on-device fetal plane classification, while slower models may remain suitable for less time-sensitive review workflows.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_108.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=vMIDFH5iYM
BibTex
@InProceedings{BarJua_DeploymentOriented_MICCAISAT2026,
author = { Barrientos, Juan AND Pérez, Michaelle AND Reyna, Favio AND Illescas, Juan AND Fajardo, Julio AND Lara, Andrea},
title = { { Deployment-Oriented Benchmarking of Optimized Deep Learning Models for Real-Time Fetal Plane Classification on Mobile Devices } },
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}
}