Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

High School Student with a Traditionally Underrepresented Background

HS Poster

Poster 48 - DL-TTR MRI: A deep learning-assisted MRI technology for simultaneous temperature mapping and tumor response monitoring

Saturday, October 26, 2024
10:00 AM - 12:00 PM EST
Location: Exhibit Hall E, F & G

Introduction: Chemical exchange saturation transfer (CEST) MRI is a new advanced MRI technology that can be used for monitoring tumor cell responses1 and temperature changes2. CEST MR images of tumors reflect cellular metabolism and energy status by the CEST signals of cellular amide (3.5 ppm), creatine (1.8 ppm), and glutamate (3 ppm) protons have been reported as imaging surrogate imaging markers. However, the broad application of CEST MRI in cancer therapy is hindered by time-consuming data acquisition and processing and data complexity. Inspired by recent revolutionary developments in deep learning-assisted MRI, we rationalize that we can leverage deep learning to map the change in tumor temperature during hyperthermal therapy accurately and, more importantly, at the same time, monitor the tumor response to the treatment. Our objective is to develop a deep learning-assisted CEST MRI as a powerful tool for precise, effective cancer treatment monitoring. This new method is called DL-TTR MRI, which stands for deep learning-assisted temperature and tumor response monitoring MRI.

Materials and

Methods: The datasets used in the present were from previously published studies, including phantom and animal data from our previous temperature MRI studies2 and animal data from tumor imaging studies 3-4. All data contains both a Z-spectrum and Water Saturation Shift Referencing (WASSR) spectrum for B0 mapping5. Temperature mapping was accomplished by directly measuring the temperature-induced water resonance (B0) shift derived from either WASSR (ground truth) 6 or Z-spectral data.
Figure 1 shows the network structure for the DL-based module in DL-B0CEST-HS, which is derived from the vanilla residual network by replacing the residual blocks with the WDSR blocks. The first layer consists of 32 convolutional filters with a 3×3 kernel size for each input image. The optimal number of WDSR blocks (8), the number of filters (32) and the expansion ratio (4×) for WDSR filters according to our previously published study7. The two separate models take the same input CEST-weighted images at different positive saturation offsets (the Z-spectrum values) to predict the CEST images associated with the offset frequencies of ±3 ppm. Patches of the initially interpolated images were used as input for both networks to expand the training samples. In the initial study, eight offsets (±100, ±200, ±300, -400, and -500Hz) from the Z-spectra were used for B0 mapping. A total of 8 sets of CEST images from 4 mice were used as training data and 4 sets of CEST images from 2 mice were used as testing data.

Results, Conclusions, and Discussions: Results and

Discussion: Both voxel- and patch-based deep neural networks were investigated. Both voxel- and patch-based deep neural networks were investigated. The voxel-based model is based on fully connected deep neuro networks containing 3 hidden layers with 100 filters. The image patch-based model is composed of a deep residual network armed with wide activation layers. The developed DL-TTR method could provide accurate B0/temperature mapping with less than 10% error using a significantly reduced acquisition time (20% total acquisition time). Moreover, we also evaluated the DL-TTR method's ability to predict tumor responses using the global features of the entire Z-spectrum in the frequency range of -4 to +4 ppm. Our results show that the CEST MRI signals at various offsets altered in response to hyperthermal therapy, where the temperature rose for an average of 3-5 oC. The trained DL-TTR method provides CEST maps similar to those of the conventional CEST MRI approach.

Conclusion: The DL-TTR method can provide simultaneous B0/temperature maps and tumor assessment using CEST data. Our initial test shows the feasibility of monitoring the tumor response to a hyperthermal treatment. Further optimization using more animal training datasets is ongoing to improve efficiency and accuracy.

Acknowledgements (Optional):

References (Optional): Refs: 1). Chan, K., et al., NMR Biomed. 2016, 29 (6), 806-16; 2). Liu, G., et al., NMR Biomed 2014, 27 (3), 320-31; 3). Han, Z., et al., NMR Biomed 2022, 35 (3), e4649; 4). Chen, H., et al., MRM 2019, 82 (4), 1471-9; 5). Kim, M., et al., MRM 2009, 61 (6), 1441-50; 6). Smith, S., et al., Magn Reson Med 2009, 62 (2), 384-93; 7). Li, Y., et al., MRM 2020, 84 (4), 1724-33.