Image pre-processing impact on generative model performance for Unsupervised Venous Malformation Segmentation

Venous malformations (VM), commonly observed in PIK3CA-related overgrowth spectrum (PROS), are infiltrative and widely distributed lesions. Repositioning of targeted therapy has recently been proposed to treat this condition, requiring an accurate volumetric quantification of these vascular lesions for monitoring treatment efficacy and adjusting dosages accordingly. So far, these malformations have only been coarsely estimated using thresholding techniques on MRI acquisitions. However, such thresholding strategies are poorly reproducible and still require manual removal of all liquid physiological structures that are erroneously over-segmented.

In this work, we developed and compared several unsupervised approaches based on reconstruction error to generate a pre-segmentation mask of VMs on mice whole-body MRI scans. Investigated deep models were trained on whole-body MRI scans of healthy mice (n=36) and evaluated on a MRI test set of PIK3CA-mutated mice (n=5). The performance of the tested models - autoencoders, generative adversarial networks (GANs), and diffusion models (DDPM) - are compared in terms of F1-score (Dice), Precision and Recall. The impact of three pre-processings onto the results is also investigated.

While baseline segmentation obtained with Otsu thresholding can reach high Dice by widely over-segmenting the MRI scans, both metrics variability and over-segmentation were improved using deep models revealing a better generalization ability. The best trade-off performance was obtained with DDPM model when using background removal pre-processing (Dice 0.50 ± 0.03) and the GAN trained on edge maps (Dice 0.47 ± 0.04).

Our results demonstrate the importance of using edge maps as input to the GAN model, and the superiority of lesion masks obtained from the diffusion model for clinical applications.