In this study, we applied for the first time a multi-contrast, multi-atlas method for automatic DTI segmentation combined with the AdaBoost classifier to classify LLD and HC. Using the trace index, the classifier reached a classification balanced accuracy of 71% and a power (i.e., ROC-AUC) of 81%, while using the RD index the classifier reached a balanced accuracy of 70% and a power of 80%. The classifier was tested for statistical significance using standard permutation with cross validation using the ROC-AUC scores and was found to classify significantly better than chance using trace and RD (permutation-based p < 0.05, for both indices).
A set of WM structures was found to be important in the classification trace and RD suggesting that LLD is characterized by a widespread axonal injury (trace, RD) and/or demyelination (RD) in limbic (fornix, uncinate fasciculus, hypothalamus), frontopontine (internal capsule, cerebral penducle), thalamo-cortical projection fibers (thalamus), fronto-striatal (caudate, external capsule), commissural fibers (corpus callosum), subcortical nuclei (substantia nigra, midbrain), brainstem and the cerebellum.
The literature on ML and DTI in LLD is limited. Patel et al. (2015) used multimodal MRI data and the Alternating Decision Tree algorithm (an ensemble classifier, similar to Adaboost) to classify 33 LLD and 35 non-depressed individuals and reported an accuracy of 87.3%. The authors suggest that global imaging measures (atrophy and global WM hyperintensity load) and non-imaging features (age and MMSE) are best predictors of diagnosis. In the study of Stolicyn et al. (2020) with 40 LLD cases and 40 controls using average FA and MD measures extracted for 19 bilateral and 5 unilateral tracts derived by TBSS and three classification models, the best classification accuracy achieved was 61.25% with MD features and the SVM classifier with optimized hyper parameters. Our study has focused on machine learning classification from an advanced DTI segmentation and the attained accuracy was 76% using both trace and RD indices (see Table 2). Our results compare to accuracies reported in the recent review on ML classification in major depression using DWI measures, where they vary from 57–91.7% (Gao et al., 2018).
Most of the studies in LLD with DTI in 3 Tesla have used voxel-based analyses (e.g. TBSS), Tractography, and ROI methods, and have mainly focused on differences between groups and in specific indices (i.e., FA and MD) (Rashidi-Ranjbar et al., 2020; Wen et al., 2014). Each of these methods carries drawbacks, such as operational burden, variability and error in manual ROI placement, fiber crossings in deterministic and complexity in probabilistic Tractography, as well as challenging investigation of the peripheral WM in voxel-based analysis. Furthermore, many predictions based on MRI variables have been made by univariate measures which reveal a moderate effect (Winter et al., 2022). The segmentation framework utilized in our study permits high registration accuracy and accurate segmentations of the superficial WM, an area that is difficult to appreciate if population-averaged atlases are used (Oishi et al., 2009) as in voxel wise DTI analyses. In our analysis, we moved from a voxel-by-voxel type of analysis, where each of the hundreds of thousands of voxels is tested individually (lowering the statistical power) to a structure-by-structure one, with only 146 anatomically relevant imaging structures covering the whole brain WM and trained an ensemble classifier for diagnostic classification.
We found widespread diffusivity alterations within various anatomical structures as important for LLD diagnosis, and fornix was the most important structure. Based on MRI studies, many underlying circuits have been proposed to be pivotal in LLD, yet direct mechanistic links are missing. Our findings are in accordance with previous studies. Specifically, limbic and frontal-subcortical circuitry disruption have been hypothesized in LLD (Alexopoulos, 2002; Phillips et al., 2003). Furthermore, brainstem nuclei have been involved in LLD (Smith et al., 2021) and this is supported by pathological findings of neuronal loss in brainstem nuclei (e.g. raphe nucleus) and presence of Lewy bodies in subcortical nuclei (e.g. substantia nigra) (Tsopelas et al., 2011; Wilson et al., 2016). Reduced FA and increased RD in the fronto-subcortical and limbic tracts (i.e. fornix and uncinate fasciculus) superior longitudinal fasciculus, and corpus callosum have been previously reported in LLD (Sexton et al., 2012). Another study found that MD was found to be increased in the fornix of patients with LLD compared to controls (Li et al., 2014). In a large sample from the UK Biobank Imaging Study, MD in anterior thalamic radiation, inferior fronto-occipital fasciculus, uncinate fasciculus, superior thalamic radiation, cingulate gyrus part of cingulum, and middle cerebellar peduncle has been associated with depressive symptoms in older individuals (Shen et al., 2019). In an analysis on Alzheimer’s disease Neuroimaging Initiative data, the presence of subclinical depressive symptoms was associated with lower WM integrity mainly in the fornix, posterior cingulum, corpus callosum and inferior longitudinal fasciculus (Touron et al., 2022). Another study demonstrated that increased anatomical connectivity predominantly in a fronto-limbic network, defined by DTI probabilistic tractography predicted depression with 91.7% accuracy using support vector machines (Fang et al., 2012). WM structures associated with subcortical gray matter nuclei (i.e., thalamus, caudate) insula and precuneus were found to be important in our study, which is in line with other studies. In particular, thalamic volume reductions were found to be significant in the meta-analysis of MRI studies in LLD (Sexton et al., 2013). Similarly, caudate nucleus (Butters et al., 2009; Kumar et al., 2004) and insula (Laird et al., 2019) volumes were found to be significantly lower in LLD. From a functional connectivity (FC) perspective, in the study of Lin et al. (2023) a diagnostic accuracy over 85% was achieved with the superior frontal gyrus, left insula, and right middle occipital gyrus using resting state (rs) fMRI and convolutional neural networks analysis. Increased right anterior insula-right dorsolateral prefrontal cortex rs-FC (Yuen et al., 2014), as well as altered fronto-cerebellar connectivity (Alalade et al., 2011) have been reported in older depressed adults with apathy. Another study found an increased FC of the left precuneus in patients with LLD compared to controls (Alexopoulos et al. 2013).
Our study has the limitations of small sample and large number of independent variables. However, Adaboost inherently performs a soft feature selection by including only those features in the model that are increasing its predictive ability. Moreover, the classifier has shown substantial improvement in the classification performance in atlas-based analyses (Zang et al., 2021). Another limitation is that the model was not tested in an independent sample. To control this we used cross validation testing the classifier on a subsample not used during the training; we also performed a permutation test to assess the significance of the model. Another limitation is that the patients were medicated.