Treatment-aware Diffusion Probabilistic Model for Longitudinal MRI Generation and Diffuse Glioma Growth Prediction

TL;DR

TaDiff addresses the challenge of predicting diffuse glioma evolution under treatment by generating future multi-parametric MRIs and tumor masks with uncertainty. The method conditions a diffusion model on sequences of past MRIs and paired treatment-day information, coupling diffusion and segmentation via joint learning and a dilated longitudinal fusion weighting. On local and external datasets, TaDiff achieves high-quality MRI generation (SSIM ≈ 0.92, PSNR ≈ 28) and accurate future tumor segmentation (DSC ≈ 0.72) while providing uncertainty maps, supporting treatment-aware planning. Limitations include data scarcity and atypical cases (e.g., second surgeries or secondary glioblastomas); future work will focus on faster sampling, more treatment types, and expanded multi-site validation to enhance clinical applicability.

Abstract

Diffuse gliomas are malignant brain tumors that grow widespread through the brain. The complex interactions between neoplastic cells and normal tissue, as well as the treatment-induced changes often encountered, make glioma tumor growth modeling challenging. In this paper, we present a novel end-to-end network capable of future predictions of tumor masks and multi-parametric magnetic resonance images (MRI) of how the tumor will look at any future time points for different treatment plans. Our approach is based on cutting-edge diffusion probabilistic models and deep-segmentation neural networks. We included sequential multi-parametric MRI and treatment information as conditioning inputs to guide the generative diffusion process as well as a joint segmentation process. This allows for tumor growth estimates and realistic MRI generation at any given treatment and time point. We trained the model using real-world postoperative longitudinal MRI data with glioma tumor growth trajectories represented as tumor segmentation maps over time. The model demonstrates promising performance across various tasks, including generating high-quality multi-parametric MRI with tumor masks, performing time-series tumor segmentations, and providing uncertainty estimates. Combined with the treatment-aware generated MRI, the tumor growth predictions with uncertainty estimates can provide useful information for clinical decision-making.

Figures (10)

• Figure 1: An overview of the TaDiff model (short for Treatment-aware Diffusion Probabilistic model). The goal of our method is to generate a set of synthetic MRIs and tumor progression masks for any given target/future treatment (e.g., TMZ: temozolomide) and time point (e.g., Day: 225) with source sequential MRIs (e.g., s1, s2, and s3) and treatments (e.g., CRT: chemoradiation at Day 36, TMZ at Days 64 and 127). More details are presented in Section \ref{['method']}.
• Figure 2: TaDiff model end-to-end pipeline for multi-parametric MRI generation and tumor growth prediction with respect to any given treatment information and target/future time point. In general, the model takes both conditioning source MRI sequences and treatment-aware embeddings as input, outputs target/future MRIs, and predicts tumor masks for both source treatment-day points and the given target treat-day point at the same time. Section \ref{['tadiff_net']} presents more details about the network architecture, and algorithms \ref{['alg:train']} and \ref{['alg:sample']} show details for training and inference with the TaDiff model.
• Figure 3: Examples of qualitative predictions on the local test cases. From top to bottom, the Present (source) MRIs with specific treatment-day traces, the Target (future) treatment-day, the Predictions (including generated MRIs, tumor masks, and uncertainty maps), and the Ground truth. The method can model both stable and progressive tumors with growth and shrinkage in different spots. Note that P-1 had a second surgery treatment which was beyond the range of treatment types our model was designed to handle.
• Figure 4: The three split violin plots compare MRI generation metric's distributions of each treatment (CRT and TMZ) overall patients. Dashed lines represent the quartiles for each group. Notice that TMZ has a long-tail distribution below the first quartile for patients P-1 and P-2 with respect to SSIM. The main reason is that P-1 was given a second surgery treatment during the TMZ period, and P-2 was diagnosed with a glioma classified as secondary glioblastoma.
• Figure 5: The two box plots show the RVD and DSC distributions for predicted future tumors and source tumors across different treatment day ranges. Note that the model's performance declined significantly with increased variability in the 221-365 day range, this is because one case (P-1) experienced unusual and rapid glioma growth and underwent a second surgery treatment during this period.