Author: Sahaja Acharya, Matthew Ladra, Junghoon Lee, Lina Mekki 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Department of Biomedical Engineering, Johns Hopkins University 🌍
Purpose: Multi-parametric MRI (mpMRI) is widely used for deep learning (DL)-based automatic segmentation of brain tumors. While multi-contrast images concatenated as channels are typically input to neural networks, relative contribution of individual input channels to the final segmentation is often unknown. We propose a method that produces segmentations computed using all input channels as well as individual channels, and associated prediction uncertainties, to assess the contribution of each input.
Methods: Pre-operative T1 contrast-enhanced (T1CE) and T2 FLAIR MR images along with manual tumor contour were collected from a retrospective cohort of 94 pediatric posterior fossa cancer patients for model training and test. The proposed framework consists of two channel-specific vision transformer encoders paired with convolutional decoders, as well as a third inter-channel convolutional decoder taking as input the encoded feature maps from each encoder. All three decoders included an evidential layer to jointly estimate the prediction uncertainty along with the segmentation. The dice similarity coefficient (DSC) was compared between all three segmentation outputs and the ground truth for 25 test cases.
Results: The proposed model achieved DSC (mean±SD) of 0.768±0.129 for T1CE-based segmentation, 0.809±0.055 for T2 FLAIR-based segmentation, and 0.842±0.051 for the multi-channel segmentation. Additionally, selecting the best-performing segmentation based on DSC for each test case resulted in a higher DSC of 0.855±0.042 compared to consistently choosing any single segmentation path, thus highlighting that a specific path did not consistently lead to the best segmentation.
Conclusion: A method to evaluate the contributions of individual input channels for mpMRI-based brain tumor segmentation was proposed. Experiments revealed that no single segmentation path consistently outperformed the others for all test cases. Future work will investigate whether the predicted uncertainties can be leveraged to automatically select the best output of the three segmentation outputs in the absence of ground truth.