Spectralligence scored at the ISBI Edited-MRS Reconstruction Challenge, April 18-21 2023, Columbia
We are pleased to announce that our MRS team has participated in the ISBI Edited MRS Reconstruction Challenge and achieved great success. Out of three categories, our team is the winner of track 1 and runner-up in tracks 2 and 3.
The ISBI Edited MRS Reconstruction Challenge presented an exciting opportunity to advance the field of edited MRS using machine learning. Edited MRS is a non-invasive technique for investigating low concentration metabolites like gamma-aminobutyric acid (GABA). However, to obtain sufficient SNR it requires multiple acquisitions resulting in long scan times. The challenge aimed to develop models that can reconstruct high-quality spectra using only a quarter of the data typically required for standard scans. The challenge consisted of three tracks: simulated data, single-vendor, and multi-vendor in-vivo data, each with edited ON and OFF spectra from GABA-edited MEGA-PRESS scans.
Our team developed a deep learning method for the reconstruction of edited MRS spectra that can operate with an arbitrary number of available measurement repetitions. The proposed approach involves computing the sample covariance matrix of the measurements and using it as the input of a convolutional neural network to extract relevant signal features and produce a high-quality spectrum. One of the most significant advantages of our method is that it can perform well with highly noisy data obtained from only a single acquisition. The method’s performance can then be further enhanced with additional acquisitions, making it versatile and efficient. This has the potential to produce immediate reconstruction during acquisitions allowing better voxel positioning, intermediate intervention, or preemptive completion of scans, which can ultimately benefit patients.
The challenge provided an excellent opportunity for us to showcase our expertise in the field of MRS and machine learning. The results of the challenge should be published on the challenge website https://sites.google.com/view/edited-mrs-rec-challenge/home soon. For those interested, we have also enclosed a 2-page manuscript that provides more information about our approach.
We are thrilled with our team’s success and excited to continue our work in this field. We look forward to sharing more developments with you in the future.
Workshop on AI for spectroscopy on April 19th 2023 at TU/e
We are thrilled to report on the success of the Workshop on AI for Spectroscopy, a collaborative event organized by the Spectralligence team. The workshop, held on April 19th, 2023, brought together experts and professionals from various institutions to explore the generalizability of AI applications across different spectral domains.
The workshop featured insightful presentations from leading companies and research organizations involved in spectroscopy, including Bruker, Spectro AG, Wageningen University, Jena University, and Avantes. These esteemed institutions shared their expertise and cutting-edge advancements in the field, sparking engaging discussions on the integration of AI and spectroscopy. In addition to the invited institutions, the workshop also featured presentations from the Spectralligence team members: Eindhoven University of Technology, Dynaxion, and Sensmet. The diverse range of perspectives and experiences contributed to a comprehensive exploration of the subject matter.
The primary objective of the workshop was to foster a dialogue centered around the generalizability of AI applications within spectroscopy and to identify common ground among researchers and industry professionals. Attendees actively participated in thought-provoking discussions, sharing insights, challenges, and best practices related to AI integration in various spectral domains.
Topics addressed during the workshop encompassed a wide range of subjects, including data acquisition, preprocessing techniques, feature selection, model development, and performance evaluation within the context of AI-driven spectroscopy. Participants explored strategies to enhance the interpretability and reliability of AI models and delved into the implications associated with their deployment in real-world scenarios.
We look forward to future workshops and initiatives that build upon the knowledge gained during this event, furthering our collective understanding and application of AI in spectroscopy.
Method for synthesizing MRS data, accepted at ISMRM 2023
We are pleased to announce that the TU/e team has proposed a new method for synthesizing single-voxel Magnetic Resonance Spectroscopy (MRS) data, which has been accepted for poster presentation at the International Society for Magnetic Resonance in Medicine (ISMRM) annual meeting of 2023.
While Deep Learning (DL) has become an increasingly useful tool for processing and analyzing MRS data, the training of such models often necessitates large datasets, which are not always available. Simulating data has been used as an alternative means of generating data, but accurately modeling macromolecules and scan imperfections has been challenging. To address this, the TU/e team proposed a conditional Variational Autoencoder (cVAE) to synthesize single-voxel MRS data. This represents a significant step towards employing deep generative modeling to enhance in-vivo datasets.
The cVAE model can generate new spectra that are similar to the original dataset, and a latent space interpolation study has shown that spectral properties (e.g., linewidth and phase shifts) can be encoded in the latent space. These early findings demonstrate the potential of deep generative modeling for MRS and pave the way for further research and development in this area. The TU/e team is excited to expand upon these promising results and explore the applications of deep generative modeling in the field of MRS and other spectroscopy domains.
Method for predicting uncertainty of metabolite quantification in MRS with applications for adaptive ensembling, accepted at ISMRM 2023
Hybrid data-driven/model-based fitting requires interpretability of neural network (NN) predictions. For that purpose, the TU/e team has developed an uncertainty measure based on the negative log-likelihood for the fitting parameters of the MRS signal model. Results indicate that the predicted uncertainties correlate well with the actual estimation errors, allowing uncertainty-based adaptive ensembling of linear combination modelling (LCM) methods and NNs. This ensembling has shown to not only outperform the individual estimators but also standard ensembling techniques. Future explorations include alternative ensembling schemes to assist LCM methods while maintaining guarantees on the certainty of concertation estimates.