Tackling Hallucination from Conditional Models for Medical Image Reconstruction with DynamicDPS

Official implementation of the MICCAI 2025 (Early Accept, Top 9%) paper

Status: Code under active development

Paper   |   Contact Author

---

Overview

Hallucinations, spurious structures not present in ground truth, pose a critical challenge in medical image reconstruction, particularly for data-driven conditional models. Our work investigates this phenomenon and introduces DynamicDPS, an innovative approach designed to mitigate hallucination while improving reconstruction fidelity and efficiency.

---

Method Overview

The schematic below illustrates our method (DynamicDPS) in comparison to traditional approaches. DynamicDPS achieves faster inference and avoids hallucination, outperforming standard conditional and diffusion models.

---

Visual Comparisons

Below: Visual comparisons on REAL low-field MR scans. DynamicDPS demonstrates superior reconstruction quality with fewer hallucinated features.

---

Usage

Note: The final cleaned-up version of the code will be released soon.

You need to unzip motionblur.zip first

Download the pretrained model here -> LINK

Training the Score-Matching Model

python image_train.py

Solve Inverse Problems (e.g., Low-Field MRI Enhancement)

python test.py

---

Citation

If you find this work useful, please consider citing:

@InProceedings{KimSeu_Tackling_MICCAI2025,
        author = { Kim, Seunghoi AND Tregidgo, Henry F. J. AND Figini, Matteo AND Jin, Chen AND Joshi, Sarang AND Alexander, Daniel C.},
        title = { { Tackling Hallucination from Conditional Models for Medical Image Reconstruction with DynamicDPS } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},
        page = {593 -- 603}
}