Vestibular schwannoma growth prediction from longitudinal MRI by time conditioned neural fields

TL;DR

Vestibular schwannoma growth prediction from longitudinal MRI is hampered by irregular scan timing and high-dimensional image data. The authors propose DeepGrowth, which encodes prior images as latent codes and represents tumors as neural-field based signed distance functions, predicting future latent codes with a time-conditioned ConvLSTM and temporal encoding to synthesize future shapes via an MLP conditioned on local codes. Training optimizes a reconstruction loss and a latent-code regularization term through a weighted objective, enabling end-to-end learning of future tumor geometries. On 131 patients with cross-validated evaluation, DeepGrowth surpasses baselines in Dice and 95% Hausdorff distance, with larger gains for the top 20% of rapidly changing tumors, highlighting the value of neural-field representations for longitudinal imaging and flexible future-time querying.

Abstract

Vestibular schwannomas (VS) are benign tumors that are generally managed by active surveillance with MRI examination. To further assist clinical decision-making and avoid overtreatment, an accurate prediction of tumor growth based on longitudinal imaging is highly desirable. In this paper, we introduce DeepGrowth, a deep learning method that incorporates neural fields and recurrent neural networks for prospective tumor growth prediction. In the proposed method, each tumor is represented as a signed distance function (SDF) conditioned on a low-dimensional latent code. Unlike previous studies that perform tumor shape prediction directly in the image space, we predict the latent codes instead and then reconstruct future shapes from it. To deal with irregular time intervals, we introduce a time-conditioned recurrent module based on a ConvLSTM and a novel temporal encoding strategy, which enables the proposed model to output varying tumor shapes over time. The experiments on an in-house longitudinal VS dataset showed that the proposed model significantly improved the performance ($\ge 1.6\%$ Dice score and $\ge0.20$ mm 95\% Hausdorff distance), in particular for top 20\% tumors that grow or shrink the most ($\ge 4.6\%$ Dice score and $\ge 0.73$ mm 95\% Hausdorff distance). Our code is available at ~\burl{https://github.com/cyjdswx/DeepGrowth}

Figures (3)

• Figure 1: The overall architecture of DeepGrowth ($N=3$). Prior scans are encoded into latent codes, which are concatenated with temporal encoding. The MLP reconstructs the future tumor as an SDF conditioned on the output of the ConvLSTM. $L_{\mathrm{rec}}$ is calculated between the predictions and SDF sampled from all three tumor masks.
• Figure 2: Example results of the different models. The first two columns are the input of the models, followed by the ground truth in the third column, and model predictions in subsequent ones. Predicted tumors are depicted in red and the ground truths in green. The dates are the study dates. The last row depicts a tumor that suddenly shrank after the second scan, which was difficult to predict for all models.
• Figure 3: Querying the proposed model at different time points (increments of 180 days). We overlaid predictions on $I_2$ for visualization in columns 3-6. The proposed model can output varied tumor shapes given different time intervals, while the model without temporal encoding outputs almost the same results regardless of the time intervals.