Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Few-Shot Fingerprinting Subject Re-Identification in 3D-MRI and 2D-X-Ray

Gonçalo Gaspar Alves, Shekoufeh Gorgi Zadeh, Andreas Husch, Ben Bausch

TL;DR

This paper tackles data leakage in multi-dataset medical imaging by introducing subject fingerprinting, a few-shot, metric-learning approach that maps all images of a subject to a coherent latent representation. Using 2D and 3D ResNet-50 backbones trained with triplet margin loss and in-batch hard negative mining, the method achieves high recall in N-way K-shot retrieval on ChestXray-14 and BraTS-2021, including challenging settings. The results demonstrate strong within-subject clustering and between-subject separation, with visualization and clustering metrics supporting the approach. This work advances cross-modal and 3D subject fingerprinting as a practical tool for detecting near-duplicates and mitigating data leakage in medical image datasets.

Abstract

Combining open-source datasets can introduce data leakage if the same subject appears in multiple sets, leading to inflated model performance. To address this, we explore subject fingerprinting, mapping all images of a subject to a distinct region in latent space, to enable subject re-identification via similarity matching. Using a ResNet-50 trained with triplet margin loss, we evaluate few-shot fingerprinting on 3D MRI and 2D X-ray data in both standard (20-way 1-shot) and challenging (1000-way 1-shot) scenarios. The model achieves high Mean- Recall-@-K scores: 99.10% (20-way 1-shot) and 90.06% (500-way 5-shot) on ChestXray-14; 99.20% (20-way 1-shot) and 98.86% (100-way 3-shot) on BraTS- 2021.

Few-Shot Fingerprinting Subject Re-Identification in 3D-MRI and 2D-X-Ray

TL;DR

This paper tackles data leakage in multi-dataset medical imaging by introducing subject fingerprinting, a few-shot, metric-learning approach that maps all images of a subject to a coherent latent representation. Using 2D and 3D ResNet-50 backbones trained with triplet margin loss and in-batch hard negative mining, the method achieves high recall in N-way K-shot retrieval on ChestXray-14 and BraTS-2021, including challenging settings. The results demonstrate strong within-subject clustering and between-subject separation, with visualization and clustering metrics supporting the approach. This work advances cross-modal and 3D subject fingerprinting as a practical tool for detecting near-duplicates and mitigating data leakage in medical image datasets.

Abstract

Combining open-source datasets can introduce data leakage if the same subject appears in multiple sets, leading to inflated model performance. To address this, we explore subject fingerprinting, mapping all images of a subject to a distinct region in latent space, to enable subject re-identification via similarity matching. Using a ResNet-50 trained with triplet margin loss, we evaluate few-shot fingerprinting on 3D MRI and 2D X-ray data in both standard (20-way 1-shot) and challenging (1000-way 1-shot) scenarios. The model achieves high Mean- Recall-@-K scores: 99.10% (20-way 1-shot) and 90.06% (500-way 5-shot) on ChestXray-14; 99.20% (20-way 1-shot) and 98.86% (100-way 3-shot) on BraTS- 2021.

Paper Structure

This paper contains 12 sections, 5 equations, 2 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Few-Shot Learning
  3. Materials and Methods
  4. Models
  5. Datasets
  6. Training
  7. Triplet Margin Loss
  8. Evaluation and Metrics
  9. Results
  10. Discussion
  11. Code and Data
  12. Acknowledgments

Figures (2)

  • Figure 1.1: T-SNE plots of 128-dimensional image embeddings from the ChestXray-14 (A) and the BraTS-2021 (B) test sets, highlighting a selection of subjects. (A & B left) Untrained ResNet-50 encodings, showing no clustering of subjects. (A & B right) ResNet-50 trained with triplet margin loss, showing clustered and well-separated subjects in latent space.
  • Figure 1.2: (A) Abstraction of the MIASD and MIESD. B) Binned Intra-Subject (Blue) and Inter-Subject Distances (Red) for all subjects in the ChestXray-14 dataset.