Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MedSymmFlow: Bridging Generative Modeling and Classification in Medical Imaging through Symmetrical Flow Matching

Francisco Caetano, Lemar Abdi, Christiaan Viviers, Amaan Valiuddin, Fons van der Sommen

TL;DR

MedSymmFlow introduces a generative-discriminative hybrid for medical imaging by extending Symmetrical Flow Matching with semantic RGB conditioning and a latent-space formulation to enable high-resolution classification, generation, and uncertainty quantification. The approach yields competitive accuracy and AUC on four MedMNISTv2 datasets while providing calibrated uncertainty through its generative sampling process and a distance-based proxy. RGB-conditioned multiclass handling and latent-space processing address scaling and discriminative limitations, supporting reliable selective prediction in clinical contexts. The work demonstrates the practicality of a unified framework that jointly learns diagnostic representations and realistic syntheses, with potential to improve data augmentation, uncertainty-aware decision-making, and safe deployment in high-stakes medical imaging applications.

Abstract

Reliable medical image classification requires accurate predictions and well-calibrated uncertainty estimates, especially in high-stakes clinical settings. This work presents MedSymmFlow, a generative-discriminative hybrid model built on Symmetrical Flow Matching, designed to unify classification, generation, and uncertainty quantification in medical imaging. MedSymmFlow leverages a latent-space formulation that scales to high-resolution inputs and introduces a semantic mask conditioning mechanism to enhance diagnostic relevance. Unlike standard discriminative models, it naturally estimates uncertainty through its generative sampling process. The model is evaluated on four MedMNIST datasets, covering a range of modalities and pathologies. The results show that MedSymmFlow matches or exceeds the performance of established baselines in classification accuracy and AUC, while also delivering reliable uncertainty estimates validated by performance improvements under selective prediction.

MedSymmFlow: Bridging Generative Modeling and Classification in Medical Imaging through Symmetrical Flow Matching

TL;DR

MedSymmFlow introduces a generative-discriminative hybrid for medical imaging by extending Symmetrical Flow Matching with semantic RGB conditioning and a latent-space formulation to enable high-resolution classification, generation, and uncertainty quantification. The approach yields competitive accuracy and AUC on four MedMNISTv2 datasets while providing calibrated uncertainty through its generative sampling process and a distance-based proxy. RGB-conditioned multiclass handling and latent-space processing address scaling and discriminative limitations, supporting reliable selective prediction in clinical contexts. The work demonstrates the practicality of a unified framework that jointly learns diagnostic representations and realistic syntheses, with potential to improve data augmentation, uncertainty-aware decision-making, and safe deployment in high-stakes medical imaging applications.

Abstract

Reliable medical image classification requires accurate predictions and well-calibrated uncertainty estimates, especially in high-stakes clinical settings. This work presents MedSymmFlow, a generative-discriminative hybrid model built on Symmetrical Flow Matching, designed to unify classification, generation, and uncertainty quantification in medical imaging. MedSymmFlow leverages a latent-space formulation that scales to high-resolution inputs and introduces a semantic mask conditioning mechanism to enhance diagnostic relevance. Unlike standard discriminative models, it naturally estimates uncertainty through its generative sampling process. The model is evaluated on four MedMNIST datasets, covering a range of modalities and pathologies. The results show that MedSymmFlow matches or exceeds the performance of established baselines in classification accuracy and AUC, while also delivering reliable uncertainty estimates validated by performance improvements under selective prediction.

Paper Structure

This paper contains 26 sections, 2 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Flow Matching
  4. Diffusion Models for Medical Imaging
  5. Generative Classifiers
  6. MedSymmFlow
  7. Semantic Conditioning via RGB Masks
  8. Latent-Space Implementation
  9. Methodology
  10. Datasets
  11. Selected Models
  12. Metrics
  13. Uncertainty Quantification
  14. Results & Discussion
  15. Quantitative Results
  16. ...and 11 more sections

Figures (3)

  • Figure 1: Overview of the MedSymmFlow architecture. The top row illustrates the image generation process, where samples evolve from Gaussian noise $x_0$ to a realistic image $x_1$ via forward integration of the conditional vector field. The bottom row shows the label prediction process, in which semantic class labels $y_1$ are transformed into predictive estimates $y_0$ through reverse-time integration.
  • Figure 2: Filtering uncertain predictions improves performance, indicating the model is well-calibrated in its confidence. However, excessive filtering eventually removes confident, correct predictions too, leading to diminished results.
  • Figure 3: Samples (224$\times$224 pixels) generated by the latent models trained on the four selected MedMNIST datasets, using an Euler ODE solver with 25 steps.