BLUE RIBBON POSTER IMAGING: Enhancing Multi Parametric MRI Glioma Subregion Segmentation By Aggregating Neural ODE Evolution Trajectories with Convlstm
Abstract
Purpose
Our previous work proposed a Neural ODE–based U-Net (NODE-UNet) that generates continuous trajectories to visualize the evolution of feature representations from the initial input to the terminal state. We hypothesize that modeling contextual consistency along these trajectories enables the capture of cross-stage feature dependencies, thereby refining segmentation in complex tumor boundary regions. This study aims to develop a trajectory-aware framework for glioma subregion segmentation using multi-parametric MRI (MP-MRI).
Methods
MP-MRI scans (T1, T1-Ce, T2, and FLAIR) from 369 glioma patients in the BraTS 2020 dataset were included. A pre-trained NODE-UNet was used to generate continuous feature evolution trajectories for each patient. Intermediate segmentation frames were sampled uniformly between the initial input state and the terminal representation state to form learnable spatial sequences. A ConvLSTM network was designed to emphasize the contextual consistency in these sequences. The ConvLSTM leverages recurrent memory mechanisms to aggregate cross-stage dependencies while preserving the spatial topology of tumors. Segmentation performance was evaluated using Dice similarity coefficient, accuracy, sensitivity, and specificity for enhancing tumor (ET), tumor core (TC), and whole tumor (WT).
Results
The proposed ConvLSTM framework achieved consistent segmentation performance across all glioma tumor subregions, with Dice coefficients of 0.7489, 0.7662, and 0.8925 for enhancing tumor (ET), tumor core (TC), and whole tumor (WT), respectively. Across all three tasks, high accuracy, sensitivity, and specificity were observed, indicating reliable voxel-level classification performance. Slice-level evaluation demonstrated stable segmentation results across tumor-containing slices, and representative visualizations confirmed accurate tumor delineation across MRI modalities, including low-contrast cases.
Conclusion
This study introduced a trajectory-aware framework that leverages NODE-generated feature evolution trajectories and a ConvLSTM module to capture contextual feature dependencies. The proposed method successfully segmented glioma subregions using MP-MRI images and shows potential for extension to other medical image analysis tasks involving complex anatomical boundaries.