2020 Fiscal Year Research-status Report
Development of Data-driven Prediction Model using 3D Multimodal Deep Neural Networks for Estimating the Evolution of White Matter Hyperintensities Associated with Small Vessel Disease in Brain MRI
Project/Area Number |
20K23356
|
Research Institution | Institute of Physical and Chemical Research |
Principal Investigator |
Rachmadi Muhammad 国立研究開発法人理化学研究所, 脳神経科学研究センター, 研究員 (60874881)
|
Project Period (FY) |
2020-09-11 – 2022-03-31
|
Keywords | White matter lesions / Progression of WMLs / Disease prediction model / Deep learning |
Outline of Annual Research Achievements |
In the last fiscal year, we focused on developing supervised deep learning model for predicting the progression of white matter hyperintensities (WMHs). Based on our previous study, the biggest challenge is to correctly segment growing and shrinking WMHs. We investigated several different modifications of U-Net to improve our last study on the same topic, and we have successfully produced some good results by using multi-head segmentation layers and generative adversarial training. Our latest result shows that our current approach successfully improved segmentation of growing and shrinking WMHs up to 0.3 Dice similarity coefficient (DSC) from 0.1 DSC in our previous study.
|
Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
We have successfully produced good results that we can publish in an international conference/workshop.
|
Strategy for Future Research Activity |
Firstly, we plan to write and publish the latest results in an international conference/workshop this summer (June/July 2021). Secondly, we plan to develop a new model based on Bayesian network and combine it with our latest model to estimate uncertainties in the model. While we have successfully produced better segmentation for growing and shrinking WMHs (from 0.1 DSC in our previous study to 0.3 DSC in this study), it is still not ideal for clinical usage. We believe that a model that can estimate uncertainties has a higher chance to be used in clinical setting. We plan to write and publish a journal paper for the completion of this project.
|
Causes of Carryover |
The funding will be used to participate in international conference/workshop related to the project (i.e., medical image computation and machine learning).
|