Abstract
Objective
DW-MRI is a promising tool for evaluation of tumor therapy response due to its ability to detect changes in local diffusion that occurs as a result of cellular damage. With the implementation of the hybrid MRILinear accelerator (MR-Linac) machine it has become feasible to collect this information during the entire radiotherapy (RT) course. This enables collection of longitudinal measurements and thus provides a conceptually new opportunity in the search for biomarkers. The aim of this work is to propose a processing framework for extraction of prognostic information from longitudinal DW-MRI data using a novel decomposition technique, and test this for stratification of RT response.
Method
In a prospective pilot study, DW-MRI brain scans were acquired on each day of a 10 fraction (total dose 30 Gy) whole brain RT regimen with a 1 T MRI scanner for a cohort of 16 patients with 31 metastases in total. The developed framework consists of an initial decomposition of the data using an extension of the non-negative matrix factorization (NMF), the monotonous slope (ms)NMF. Second, a transformation from voxel-maps to tumor-specific features using descriptive statistics, and a robust fit across days to capture temporal changes of these. Finally, a logistic regression with an integrated feature selection in a stratified k-fold cross-validation process was used for tumor stratification. A follow-up scan 2-3 mos post treatment was used for true labelling of the tumors (8 non-responders, 22 responders). The performance was reported using the balanced accuracy (due to class imbalance) and the AUC.
Results
The balanced accuracy of the prediction was 69 % and the AUC was 0.7.
Conclusion
A novel decomposition-based processing framework for prediction of radiotherapy response from longitudinal DW-MRI measurements was demonstrated in a small patient cohort. Large DW-MRI data sets are easily acquired with the MR-Linac and are expected to be used in further testing of the framework.
DW-MRI is a promising tool for evaluation of tumor therapy response due to its ability to detect changes in local diffusion that occurs as a result of cellular damage. With the implementation of the hybrid MRILinear accelerator (MR-Linac) machine it has become feasible to collect this information during the entire radiotherapy (RT) course. This enables collection of longitudinal measurements and thus provides a conceptually new opportunity in the search for biomarkers. The aim of this work is to propose a processing framework for extraction of prognostic information from longitudinal DW-MRI data using a novel decomposition technique, and test this for stratification of RT response.
Method
In a prospective pilot study, DW-MRI brain scans were acquired on each day of a 10 fraction (total dose 30 Gy) whole brain RT regimen with a 1 T MRI scanner for a cohort of 16 patients with 31 metastases in total. The developed framework consists of an initial decomposition of the data using an extension of the non-negative matrix factorization (NMF), the monotonous slope (ms)NMF. Second, a transformation from voxel-maps to tumor-specific features using descriptive statistics, and a robust fit across days to capture temporal changes of these. Finally, a logistic regression with an integrated feature selection in a stratified k-fold cross-validation process was used for tumor stratification. A follow-up scan 2-3 mos post treatment was used for true labelling of the tumors (8 non-responders, 22 responders). The performance was reported using the balanced accuracy (due to class imbalance) and the AUC.
Results
The balanced accuracy of the prediction was 69 % and the AUC was 0.7.
Conclusion
A novel decomposition-based processing framework for prediction of radiotherapy response from longitudinal DW-MRI measurements was demonstrated in a small patient cohort. Large DW-MRI data sets are easily acquired with the MR-Linac and are expected to be used in further testing of the framework.
| Original language | English |
|---|---|
| Publication date | 2021 |
| Publication status | Published - 2021 |
| Event | Virtual 8th MR in RT Symposium - Online Duration: 19 Apr 2021 → 21 Apr 2021 Conference number: 8 |
Conference
| Conference | Virtual 8th MR in RT Symposium |
|---|---|
| Number | 8 |
| Location | Online |
| Period | 19/04/2021 → 21/04/2021 |