A Deep Learning Framework for Institution-Independent PTV Delineation In Glioblastoma Using Heterogeneous MRI Datasets
Abstract
Purpose
To develop a robust, institution-independent quality assurance model using heterogeneous datasets. By prioritizing generalizability over site-specific tuning, this model aims to predict and assist in the delineation of PTV contours for glioblastoma (GBM) using MRI T1-CE and T2-FLAIR imaging.
Methods
The Brain tumor segmentation (BRaTS) 2021 challenge made public a set of 1251 anonymized glioblastoma MRI data including T1CE and Flair MRI images which include segmentations of Edema (Ed), Necrotic/Non-enhancing tumor Core (NET) and GD- enhancing tumor (GET). Using this dataset, a MONAI based dynUnet framework was used to create a deep learning algorithm. A training, testing and validation split of (70%, 15%,15%) and a combined Dice and focal loss-based metric were utilized to train the model. The BraTS model was validated using the Burdenko GBM Progression dataset. Utilizing T1CE and FLAIR sequences, the model again predicted Ed, NET, and GET segments, which were combined into a single 'predicted Whole Tumor' (pWT) volume. Due to limitations of Burdenko’s dataset only PTV information was available and thus a 15mm expansion was estimated and applied to our pWT volume based on ESTRO contouring guidelines creating our evaluation volume termed PTV_QA.
Results
Tumor delineation achieved a median/mean/IQR as follows: Overall Dice {.83, .79, .89 - .71), Dice excluding NET: {.85, .81, .90 - .76}, 95% Housdorff Distance(mm) {2.4, 3.9, 5.14 – 1.48 } and Average symmetric distance (mm) {.54, .92, 1.19 - .38}. In terms of PTV agreement with the PTV_QA the average inclusion ratio was {94.3%}, with the Median/Mean Dice {.79, .75}.
Conclusion
These results demonstrate the feasibility of a novel validation metric to help mitigate institutional biases and variations. By using multi-institutional datasets, we were able to create a model which is nationally cross validated. These results support the planned implementation of local data into the model for in-house use.