Session Item

Magnetic Resonance Imaging (MRI) is being routinely used for treatment planning in MR-guided radiotherapy, however, the sequences available in MR-Linacs may not be perfectly optimized in terms of contrast and noise, which can facilitate tumor and organ-at-risk delineation, registration and synthetic CT calculation. These parameters can be optimized by a variety of MR sequence parameter sets (SPS), which directly affect the image quality in terms of signal-(SNR) or contrast-to-noise ratio (CNR). Depending on the sequence and clinical objective, these SPS can include up to 30 individual parameters. This work aims to develop a software tool for the optimization of SNR and CNR in MRI sequences based on the applied SPS. Here we present the preliminary results of the evaluation of two different regression techniques for the SNR and CNR prediction.

Initially, two different models to predict the quality parameters (SNR/CNR), depending on the applied SPS, were investigated and trained. Training data sets were acquired at a 1.5 T MRI (Aera, Siemens) with a dedicated phantom with in-house fabricated anthropomorphic contrast inserts of different concentration of agarose and Ni-DTPA. Measurements were performed with a turbo spin echo sequence with spatial resolution of 0.4 x 0.4 x 5 mm³, Bandwidth 186 Hz/Pixel and 4 different, varying parameters (repetition time (TR), echo time (TE), turbo factor (TF) and flip angle (FA)) to generate a total of 1114 different SPS-combinations. The models used for regression were a deep learning (DL) based method with five hidden layers with Relu activation function and one input and output layer, and a generalized additive model (GAM) based on spline functions. The models were evaluated on training (90% of the total data set) and test datasets (10% of the total data set) with two different standard loss functions for mean absolute error (MAE) and mean square error (MSE).

The comparison of the two different models (Figure 1, Table 1) shows that the DL based model yields a higher accuracy (MSE and MAE) compared to the GAM for the test data. While the GAM performs well on smaller data sets and the training data, the DL-based model outperforms the GAM for larger training data sets and the subsequent application on the test data set. 

As a first step towards the development of an optimization tool for MR sequences, the DL-based model maintained a higher prediction accuracy for both SNR and CNR in the validation data set. The next step is to develop and implement multi-objective optimization methods with regards to SNR and CNR based regression.