Research

Paper

AI LLM March 05, 2026

MobileFetalCLIP: Selective Repulsive Knowledge Distillation for Mobile Fetal Ultrasound Analysis

Authors

Numan Saeed, Fadillah Adamsyah Maani, Mohammad Yaqub

Abstract

Fetal ultrasound AI could transform prenatal care in low-resource settings, yet current foundation models exceed 300M visual parameters, precluding deployment on point-of-care devices. Standard knowledge distillation fails under such extreme capacity gaps (~26x), as compact students waste capacity mimicking architectural artifacts of oversized teachers. We introduce Selective Repulsive Knowledge Distillation, which decomposes contrastive KD into diagonal and off-diagonal components: matched pair alignment is preserved while the off-diagonal weight decays into negative values, repelling the student from the teacher's inter-class confusions and forcing discovery of architecturally native features. Our 11.4M parameter student surpasses the 304M-parameter FetalCLIP teacher on zero-shot HC18 biometry validity (88.6% vs. 83.5%) and brain sub-plane F1 (0.784 vs. 0.702), while running at 1.6 ms on iPhone 16 Pro, enabling real-time assistive AI on handheld ultrasound devices. Our code, models, and app are publicly available at https://github.com/numanai/MobileFetalCLIP.

Metadata

arXiv ID: 2603.05421
Provider: ARXIV
Primary Category: cs.CV
Published: 2026-03-05
Fetched: 2026-03-06 14:20

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.05421v1</id>\n    <title>MobileFetalCLIP: Selective Repulsive Knowledge Distillation for Mobile Fetal Ultrasound Analysis</title>\n    <updated>2026-03-05T17:43:00Z</updated>\n    <link href='https://arxiv.org/abs/2603.05421v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.05421v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Fetal ultrasound AI could transform prenatal care in low-resource settings, yet current foundation models exceed 300M visual parameters, precluding deployment on point-of-care devices. Standard knowledge distillation fails under such extreme capacity gaps (~26x), as compact students waste capacity mimicking architectural artifacts of oversized teachers. We introduce Selective Repulsive Knowledge Distillation, which decomposes contrastive KD into diagonal and off-diagonal components: matched pair alignment is preserved while the off-diagonal weight decays into negative values, repelling the student from the teacher's inter-class confusions and forcing discovery of architecturally native features. Our 11.4M parameter student surpasses the 304M-parameter FetalCLIP teacher on zero-shot HC18 biometry validity (88.6% vs. 83.5%) and brain sub-plane F1 (0.784 vs. 0.702), while running at 1.6 ms on iPhone 16 Pro, enabling real-time assistive AI on handheld ultrasound devices. Our code, models, and app are publicly available at https://github.com/numanai/MobileFetalCLIP.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CV'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.LG'/>\n    <published>2026-03-05T17:43:00Z</published>\n    <arxiv:comment>Project website: www.numansaeed.com/mobilefetalclip</arxiv:comment>\n    <arxiv:primary_category term='cs.CV'/>\n    <author>\n      <name>Numan Saeed</name>\n    </author>\n    <author>\n      <name>Fadillah Adamsyah Maani</name>\n    </author>\n    <author>\n      <name>Mohammad Yaqub</name>\n    </author>\n  </entry>"
}