Bridging Classical and Learned Priors: A Hybrid Framework for Medical Image Enhancement

Peeyush Kumar Singh, SnehaSingh

Keywords: Medical Image Enhancement, Diffusion models, Prior-guided generation, Ultrasound, MRI, Synthesis

TL;DR: By guiding diffusion models with pseudo-labels instead of direct generation, we enable training free medical image enhancement that improves downstream task performance without risking clinical hallucinations.

Abstract:

Medical image enhancement faces a fundamental trade-off: classical methods preserve anatomical fidelity but over-smooth fine structures, while deep learning approaches risk generating unrealistic artifacts on limited clinical data. We introduce a hybrid framework combining classical preprocessing with pretrained diffusion priors for high-quality enhancement across modalities.
Our method leverages pretrained Stable Diffusion model without requiring domain specific training. During inference, classical enhancement methods generate pseudo-labels. The frozen diffusion model leverages its learned priors to refine fine structures while gradient-based guidance anchors generation to the pseudo-label, preventing hallucinations. We demonstrate efficacy in ultrasound and MRI segmentation and achieve significant improvements in multi-class cardiac structure segmentation compared to baseline models. Critical insights include: pseudo-labels outperform multi-stage classical pipelines by providing differentiable guidance targets for diffusion models, testing segmentation models on enhanced images yields additional performance gains, pseudo-label guidance strength requires domain specific tuning to balance classical robustness with learned refinement.
With extensive evaluation across imaging modalities, we show that pretrained diffusion models can enhance medical images while preserving the interpretability and diagnostic fidelity essential for clinical deployment.

Primary Subject Area: Image Synthesis

Secondary Subject Area: Image Acquisition and Reconstruction

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 391