Key result
ML trained on synthetic data outperforms conventional methods in predicting APT effect in stroke.
Answers are grounded in this paper and cite their sources.
An ML approach using partially synthetic data improves the accuracy and speed of amide proton transfer imaging in an animal stroke model.
Amide proton transfer (APT) imaging, a technique sensitive to tissue pH, holds promise in the diagnosis of ischemic stroke. Achieving accurate and rapid APT imaging is crucial for this application. However, conventional APT quantification methods either lack accuracy or are time-consuming. Machine learning (ML) has recently been recognized as a potential solution to improve APT quantification. In this paper, we applied an ML model trained on a new type of partially synthetic data, along with an optimization approach utilizing recursive feature elimination, to predict APT imaging in an animal stroke model. This partially synthetic datum is not a simple blend of measured and simulated chemical exchange saturation transfer (CEST) signals. Rather, it integrates the underlying components including all CEST, direct water saturation, and magnetization transfer effects partly derived from measurements and simulations to reconstruct the CEST signals using an inverse summation relationship. Training with partially synthetic data requires less in vivo data compared to training entirely with fully synthetic or in vivo data, making it a more practical approach. Since this type of data closely resembles real tissue, it leads to more accurate predictions than ML models trained on fully synthetic data. Results indicate that an ML model trained on this partially synthetic data can successfully predict the APT effect with enhanced accuracy, providing significant contrast between stroke lesions and normal tissues, thus clearly delineating lesions. In contrast, conventional quantification methods such as the asymmetric analysis method, three-point method, and multiple-pool model Lorentzian fit showed inadequate accuracy in quantifying the APT effect. Moreover, ML methods trained using in vivo data and fully synthetic data exhibited poor predictive performance due to insufficient training data and inaccurate simulation pool settings or parameter ranges, respectively. Following optimization, only 13 frequency offsets were selected from the initial 69, resulting in significantly reduced scan time.
No takes yet. Share an insight, caveat, or question.
Viswanathan et al. (2024) studied ischemic stroke. Machine learning model trained on partially synthetic data vs. Conventional quantification methods and ML trained on in vivo or fully synthetic data was evaluated on Prediction of APT effect. A machine learning model trained on partially synthetic data successfully predicted the APT effect with enhanced accuracy compared to conventional methods, clearly delineating stroke lesions.
Synapse has enriched 10 related papers. Consider them for comparative context: