LeFusion: Controllable Pathology Synthesis via Lesion-Focused Diffusion Models
Authors: Hantao Zhang, Yuhe Liu, Jiancheng Yang, Shouhong Wan, Xinyuan Wang, Wei Peng, Pascal Fua
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Validated on 3D cardiac lesion MRI and lung nodule CT datasets, Le Fusion-generated data significantly improves the performance of state-of-the-art segmentation models, including nn UNet and Swin UNETR. Code and model are available at https://github.com/M3DV/Le Fusion. |
| Researcher Affiliation | Academia | 1Swiss Federal Institute of Technology Lausanne (EPFL), Switzerland 2University of Science and Technology of China (USTC), China 3Beihang University, China 4Stanford University, USA |
| Pseudocode | No | The paper describes the diffusion model and its training objectives using mathematical equations (Eq. 1-6) and descriptive text, but no explicit pseudocode or algorithm block is provided. |
| Open Source Code | Yes | Code and model are available at https://github.com/M3DV/Le Fusion. |
| Open Datasets | Yes | LIDC: Multi-Peak Lung Nodule CT. We use LIDC dataset (Armato III et al., 2011)... Emidec: Multi-Class Cardiac Lesion MRI. The Emidec dataset (Lalande et al., 2022) |
| Dataset Splits | Yes | LIDC: The dataset was divided into an 808-case training set, comprising 2,104 lung nodule ROIs, and a 202-case test set, containing 520 lung nodule ROIs. Additionally, 3,076 normal (N) ROIs were cropped from the 135 healthy patients... Emidec: We split the 67 P cases into 57 for training and 10 for testing. |
| Hardware Specification | Yes | For the entire experiment, we used 6*A100 (40G) GPUs |
| Software Dependencies | Yes | Python 3.8 and Py Torch version 2.4.0. |
| Experiment Setup | Yes | all diffusion models were set to 300 timesteps. For both datasets, we adopted a learning rate of 1e-4 and a batch size of 16. The training process required approximately 30,000 timesteps for the cardiac dataset and 40,000 timesteps for the LIDC lung nodule dataset. For downstream tasks, both Swin UNETR and nn UNet were trained for 200 epochs. |