Detecting microstructural deviations in individuals with deep diffusion MRI tractometry

Abstract

Most diffusion magnetic resonance imaging studies of disease rely on statistical comparisons between large groups of patients and healthy participants to infer altered tissue states in the brain; however, clinical heterogeneity can greatly challenge their discriminative power. There is currently an unmet need to move away from the current approach of group-wise comparisons to methods with the sensitivity to detect altered tissue states at the individual level. This would ultimately enable the early detection and interpretation of microstructural abnormalities in individual patients, an important step towards personalized medicine in translational imaging. To this end, Detect was developed to advance diffusion magnetic resonance imaging tractometry towards single-patient analysis. By operating on the manifold of white-matter pathways and learning normative microstructural features, our framework captures idiosyncrasies in patterns along white-matter pathways. Our approach paves the way from traditional group-based comparisons to true personalized radiology, taking microstructural imaging from the bench to the bedside.

Fig. 1. Graphical representation of the proposed anomaly detection framework.

The neural network consists of a deep autoencoder symmetrically designed with five fully connected layers. The input and output layers have exactly the same number of nodes as the number of input tracts features (|x|, represented by the colored vector). The goal of the network is to generate an output $(\widehat{x})$ similar to the input (x) by minimizing the reconstruction error (D). Here, the mean absolute error (MAE) was used as the anomaly score; MAE measures the average magnitude of the errors and is derived during testing by computing the absolute differences between the reconstructed microstructural features $\left({\widehat{x}}_i\right)$ and the raw input features (x_i,). The number of input features is denoted by n, with the current feature denoted by í. Node coloring is for illustrative purposes only.

Fig. 2. Anomaly scores for the CNV dataset.

a, The autoencoder (AE) network and PCA approaches provided better discriminating power in terms of sensitivity/specificity tradeoffs compared with traditional linear univariate approaches (*P=5.69×10^-40 and P=1.92×10^-33, Bonferroni corrected with α=0.01, two-tailed f-tests for AE-z and PCA-z, respectively). The average AUC over 100 iterations (bottom) for the RISH0 feature (top) is displayed (data are presented as mean values ± s.d.). b, The RISH0 features show higher reconstruction error for the CNV (orange box plot) compared with the typically developing patients (purple box plot) with a precision-recall AUC of 0.45 (center line, median; box limits, upper and lower quartiles; whiskers, 1.5 interquartile range; n = 90 healthy participants and n = 8 CNVs). In comparison, a random classifier would score 0.08. The box shows the quartiles of the dataset whereas the whiskers extend to show the rest of the distribution. c, From a group perspective, anomaly rates were mostly observed in the ILF (color map: RISH0, lateral view), optic radiations and SLF. OR, optic radiations; UF, uncinate fasciculus; AF, arcuate fasciculus; IFO, inferior fronto-occipital fasciculus; CC, corpus callosum; Cg, cingulum; ATR, anterior thalamic radiation, R, right brain hemisphere; L, left brain hemisphere.

Fig. 3. Along-tract anomalies in a single patient.

Reconstructed tract profiles of CNV (top) and typically developing (bottom) patients reveals RISH0 discrepancies along various association bundles such as the AF, ILF and SLF. The marked pink areas represent sections along the bundle where anomalies are detected by the leave-one-out cross-validation approach (threshold = 1 per number of healthy participants). Specifically, the autoencoder is trained using healthy participant data only, and compared with the FCD patient. Then for each of the healthy participants, the FCD patient is shuffled back in the population and the newly left out healthy participant is compared with that population. We then assess the MAE. The letters correspond to the tract list found in the Supplementary Information.

Fig. 4. Focal cortical dysplasia anomaly detection (patient 1).

a, The T2 hyperintense lesion, located at the base of the skull in the temporal lobe, is shown. b–d, Several pathways with anomalies interdigitate in the vicinity of the lesion. Although the inferior fronto-occipital fasciculus (c; IFOF with RISH0 colormap overlayed and the 20 along-tract sections underlayed) signal did not extend beyond the shaded areas (d; ±1z-score), the proposed anomaly detection framework identified abnormalities in that region (b, pink shaded areas; bold orange line, original tract-profiles; dotted purple line, reconstructed representation learnt from the network). R, right side; L, left side; A, anterior; P, posterior.

Fig. 5. FCD anomaly detection (patient 2).

a, The lesion is located anterior to the right primary motor cortex in the supplementary motor area (hyperintense signal on the FLAIR image, hypointense on the RISH map). Tractography show tracts traversing the area (CC4). b, Anomalies were identified in the right CC4, CST and SLF-I bundles (top, pink shaded areas). The bold orange line represents the original tract-profiles whereas the dotted purple line represents the reconstructed representation learned from the network. The z-score approach shows less focused anomaly patterns along the tracts (shaded area = ±1z-score).

Fig. 6. Anomaly scores for the SCHZ cohort.

a, The autoencoder provides a better discriminating power compared with traditional linear univariate and multivariate approaches with a mean AUC of 0.64 ± 0.06 (bar graph, *P=1.00×10^-33, P=1.05×10^-5 and P=1.63×10^-26, Bonferroni corrected at α = 0.01, two-tailed t-tests for AE-PCA, AE-z and PCA-z, respectively). For illustrative purposes, anomaly scores derived from the autoencoder, PCA and z-score are correlated with the Hopkins anxiety score (shaded area: 95% confidence interval). b, The RISH0 features show higher reconstruction error (100 averages) for the SCHZ than the healthy participants (t = -2.48, P = 0.01, Cohen’s d = 0.47, two-sided t-test; center line, median; box limits, upper and lower quartiles; whiskers, 1.5×interquartile range; n = 109 healthy participants and n = 43 SCZH, respectively).