Related Experiment Video
Updated: Apr 14, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Image pre-processing impact on generative model performance for Unsupervised Venous Malformation Segmentation
Antoine Fraissenon1, Alisa Kugusheva2, Sophia Ladraa3
1INSA-Lyon, Universite Claude Bernard Lyon 1, UJM-Saint Etienne, CNRS, Inserm, CREATIS UMR 5220, U1294, Lyon, 69621, France; INSERM Unité 1151, Institut Necker-Enfants Malades, Paris, 75015, France; Service d'Imagerie Pédiatrique, Centre de référence des anomalies vasculaires superficielles, Hôpital Femme-Mère-Enfant, Hospices Civils de Lyon, Bron, 69500, France; Service de Radiologie Mère-Enfant, Hôpital Nord, Saint Etienne, 42000, France.
Deep learning models, including diffusion models (DDPM) and generative adversarial networks (GANs), significantly improve the segmentation of venous malformations (VMs) in mice. These advanced methods offer superior accuracy and reproducibility compared to traditional thresholding for monitoring PROS treatments.
Area of Science:
- Medical imaging analysis
- Deep learning for biomedical applications
- Vascular malformation research
Background:
- Venous malformations (VMs) are common in PIK3CA-related overgrowth spectrum (PROS).
- Accurate volumetric quantification of VMs is crucial for monitoring targeted therapy efficacy.
- Current thresholding methods for VM segmentation on MRI are imprecise and require manual correction.
Purpose of the Study:
- To develop and compare unsupervised deep learning models for VM pre-segmentation on mice whole-body MRI.
- To evaluate the performance of autoencoders, GANs, and DDPMs for VM quantification.
- To assess the impact of pre-processing techniques on segmentation accuracy.
Main Methods:
- Trained deep learning models (autoencoders, GANs, DDPMs) on healthy mice MRI scans (n=36).
- Evaluated models on PIK3CA-mutated mice MRI scans (n=5).
- Compared segmentation performance using F1-score (Dice), Precision, and Recall, investigating pre-processing impacts.
Main Results:
- Deep learning models demonstrated improved generalization and reduced over-segmentation compared to Otsu thresholding.
- The DDPM with background removal achieved the best performance (Dice 0.50 ± 0.03).
- A GAN trained on edge maps also showed strong results (Dice 0.47 ± 0.04).
Conclusions:
- Edge map input enhances GAN model performance for VM segmentation.
- Diffusion models provide superior lesion masks for clinical applications in PROS.
- Unsupervised deep learning offers a more reproducible and accurate approach to VM quantification.