Pre to Post-Treatment Glioblastoma MRI Prediction using a Latent Diffusion Model

TL;DR

This paper addresses the problem of early prediction of glioblastoma treatment response by generating a post-treatment MRI from a pre-treatment MRI to enable earlier assessment before the standard $4$-month post imaging. The authors introduce a 2D Latent Diffusion Model operating in latent spaces learned by a pair of VQ-VAE encoders, conditioned on the pre-treatment latent $z^{\text{pre}}$ and the Gross Tumor Volume latent $z'^{\text{gtv}}$, with classifier-free guidance from a survival predictor. On a local dataset of $N=140$ GBM patients (6,059 slices), they compare 2-class and 4-class survival conditioning and report that the 2-class setup yields the best overall results, with 4-class conditioning improving local metrics; however, the model struggles to reproduce post-treatment $T1$-enhancement and lacks 3D context. The work demonstrates the feasibility of latent-diffusion-based early visual TRP for GBM and highlights avenues for improvement, including clinically aligned evaluation metrics and extending to 3D context to better inform radiotherapy planning.

Abstract

Glioblastoma (GBM) is an aggressive primary brain tumor with a median survival of approximately 15 months. In clinical practice, the Stupp protocol serves as the standard first-line treatment. However, patients exhibit highly heterogeneous therapeutic responses which required at least two months before first visual impact can be observed, typically with MRI. Early prediction treatment response is crucial for advancing personalized medicine. Disease Progression Modeling (DPM) aims to capture the trajectory of disease evolution, while Treatment Response Prediction (TRP) focuses on assessing the impact of therapeutic interventions. Whereas most TRP approaches primarly rely on timeseries data, we consider the problem of early visual TRP as a slice-to-slice translation model generating post-treatment MRI from a pre-treatment MRI, thus reflecting the tumor evolution. To address this problem we propose a Latent Diffusion Model with a concatenation-based conditioning from the pre-treatment MRI and the tumor localization, and a classifier-free guidance to enhance generation quality using survival information, in particular post-treatment tumor evolution. Our model were trained and tested on a local dataset consisting of 140 GBM patients collected at Centre François Baclesse. For each patient we collected pre and post T1-Gd MRI, tumor localization manually delineated in the pre-treatment MRI by medical experts, and survival information.

Figures (4)

• Figure 1: Inference pipeline
• Figure 2: VQ-VAE diagram with pre-treatment MRI as example
• Figure 3: Training pipeline
• Figure 4: (Best viewed in color) Post MRI predictions with the pred 2-class LDM model on the test set, ordered by descending local SSIM (ranging from 0.71 to 0.02). The first three columns are the Pre MRI with GTV in red, the predicted Post MRI, and the ground truth Post MRI. Similarly, the last three columns contain the same elements, but centered on the tumor crop.