Quantification of MR spectra by deep learning in an idealized setting: Investigation of forms of input, network architectures, optimization by ensembles of networks, and training bias

Abstract

The aims of this work are (1) to explore deep learning (DL) architectures, spectroscopic input types, and learning designs toward optimal quantification in MR spectroscopy of simulated pathological spectra; and (2) to demonstrate accuracy and precision of DL predictions in view of inherent bias toward the training distribution. Methods: Simulated 1D spectra and 2D spectrograms that mimic an extensive range of pathological in vivo conditions are used to train and test 24 different DL architectures. Active learning through altered training and testing data distributions is probed to optimize quantification performance. Ensembles of networks are explored to improve DL robustness and reduce the variance of estimates. A set of scores compares performances of DL predictions and traditional model fitting (MF).Results: Ensembles of heterogeneous networks that combine 1D frequency-domain and 2D time-frequency domain spectrograms as input perform best. Dataset augmentation with active learning can improve performance, but gains are limited. MF is more accurate, although DL appears to be more precise at low SNR. However, this overall improved precision originates from a strong bias for cases with high uncertainty toward the dataset the network has been trained with, tending toward its average value. Conclusion:MF mostly performs better compared to the faster DL approach. Potential intrinsic biases on training sets are dangerous in a clinical context that requires the algorithm to be unbiased to outliers (i.e., pathological data). Active learning and ensemble of networks are good strategies to improve prediction performances. However, data quality (sufficient SNR) has proven as a bottleneck for adequate unbiased performance-like in the case of MF.

A review of machine learning applications for the proton MR spectroscopy workflow

“…CNNs are widely adapted in computer vision and medical imaging 93,94 with their benefits of weight sharing, simultaneously extracting features and performing classification, and easy implementation into large-scale networks. 95 While recent studies from Rizzo et al 49 and Shamaei et al 51 compare different model architectures, more comparison studies are still missing. Future work should focus on testing different and new model types like transformers that have shown potential in other medical imaging fields.…”

“…Rizzo et al 49 compare MRS quantification using various CNN models, input types, and learning methods. They use a simulated artifact‐free dataset with GT concentrations to enable fair comparison with standard model fitting.…”

Denoising single MR spectra by deep learning: Miracle or mirage?

“…To account for substantial variations in metabolite content in pathological conditions, metabolite concentrations ranged from 0 to double that seen in healthy brain 24 (except for tCho up to five‐times normal). A reference signal, representing a downscaled non‐suppressed water signal supposedly obtained from a separate acquisition, was added at 0.5 ppm to facilitate quantification 25 . The following parameters were varied to mimic in vivo conditions: shim with Gaussian line‐broadening of 2–5 Hz added to the inherent line broadening of 1.1–3.9 Hz because of assumed metabolite T 2 s 26,27 (tCr [CH 2 ]: 111 ms, tCr [CH 3 ]: 169 ms, NAA [CH 3 ]: 289 ms, all other protons: 185 ms), overall SNR of the spectrum of 5 to 40 (termed global SNR and defined in time‐domain as absolute signal intensity at time 0 versus the standard deviation of the noise), and the intensity of the macro‐molecular background signal 27 at ±33% of the norm.…”

“…In a first denoising algorithm (referred to as 2D‐UNet) a time‐frequency representation (spectrogram), which has proven beneficial for audio signal processing as well as for quantification of MR spectra, 25 was selected as raw data form for denoising. In this approach, synthesized time‐domain signals were transformed into spectrograms using a short‐time Fourier transformation (“stft” in MATLAB) with window size 128 and overlap interval of 97 points converting time‐domain signals of 4096 points into 128 × 128 spectrograms.…”

“…(1) Metabolites with low and medium SNR (e.g., sI, Lac, NAAG): denoising appears to be effective because it prevents outliers and may, therefore, lead to lower mean deviations in the estimates (σ), but it strongly biases the fit results to a Gaussian distribution around the center of the training concentration range, similar to what was reported in a different context. 25,52 This trend is also reflected in the correlation coefficients that may increase with denoising, but the slope of the correlation line is decreasing-reflecting estimation bias. It should be noted…”