Room: Exhibit Hall | Forum 8
Purpose: In order to characterize tumor diffusion properties beyond the mono-exponential model, it needs to acquire diffusion weighted (DW) images at multiple b-values, which prolongs scanning time. This study aimed to develop a method based on machine learning for rapid multiple b-value diffusion imaging.
Methods: Eleven b-value (0-2500 s/mm2) DW images were acquired in 29 patients with glioblastoma. Neural network (NN) models were trained using 4 b-value DW signals to target 11 b-value ones. The NN had 4 fully connected layers. A total of 1.5x106 DW curves were extracted from T2 tumor volumes and divided into training (46%), validation (14%) and testing datasets (30%). Four inputs were S0, S1, one between S2 and S8, and one from S9 or S10 where Si is the DW intensity at the ith b-value. Two normalization schemes were compared. One allowed 2% Gaussian deviation at S0 (GN). Another scheme scaled the first input value to the averaged value of S0 and S1 (AvG) to reduce variance within the curve. Additional NN models were trained separately using 3 classes of DW curves clustered by fuzzy c-means to accommodate for variations among curves. A total of four models were compared: 1. GN data trained by a single NN model (GN-NN); 2. GN data trained by 3-cluster based NN models (GN-3CNN); 3. AvG data trained by a single NN model (AvG-NN); and 4. AvG data trained by 3-cluster based NN models (AvG-3CNN).
Results: Mean absolute percentage errors of testing were 10.1+/-6.4% for GN-NN, 15.5+/-10.5% for GN-3CNN, 9.9+/-6.1% for AvG-NN, and 10.75+/-6.2% (AvG-3CNN).
Conclusion: Method 3 produces more accurate predictions. Method 4 has a slightly worse accuracy possibly due to misclassification of DW curves. This work can be extended to whole brain. Other input selection schemes can be tested. This method has the potential to accelerate DW imaging.