Federated breast-imaging models may be confounded by acquisition year
Original title: Evaluating Federated Vision-Language and Vision-Only Models for Privacy-Aware Breast Cancer Imaging
Normal-versus-sick ROC-AUC was 0.940-0.990, but year alone scored 0.974 on the same partition.
Researchers compared CLIP, BiomedCLIP, and an ImageNet-pretrained ViT-B/16 in a federated breast-imaging setup with client-level differential privacy. All five models separated normal from sick cases well, yet acquisition year alone reached 0.974, raising a confounding concern. Benign-versus-malignant discrimination stayed uncertain, with ROC-AUC from 0.504 to 0.637 and every confidence interval including 0.50. Removing CLIP's text pathway preserved accuracy and halved communication cost, while differential privacy reduced ROC-AUC by 0.375.
Unconfirmed: The authors say high normal-versus-sick scores may reflect acquisition-year confounding rather than pathology recognition.
- Evaluation used patient-level ROC-AUC and bootstrap 95% confidence intervals.
- A CLIP ViT-B/16 ablation removed the text pathway.
- Privacy noise was applied to the full model state, not only trainable parameters.
- Two public breast imaging datasets were used.
- AI models
- CLIPBiomedCLIPViT-B/16
From the original article
1 min read at arxiv.orgArtificial intelligence (AI) shows promise for breast imaging, but limited labeled data, institutional data fragmentation, and privacy concerns hinder centralized training. We investigate a privacy-aware federated learning framework combining pretrained vision-language models (VLMs) with client-level differential privacy.
By Malithi Mithsara Wanniarachchi Kankanamge · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.