The major finding of this study was that the CNN-estimated MD values in the left SN and bilateral caudate increased in PD patients compared to HCs. Additionally, the CNN-estimated FA and MK values in the right SN negatively correlated with H&Y scales and CNN-estimated FA values in the left putamen positively correlated with H&Y scales. In contrast, with model fitting method, there was no significant difference of MD values in the SN between PD patients and HCs, and there was no significant correlation between DKI scalar measures and clinical assessments in PD patients. Our findings suggest that the CNN method have the potential to optimize the estimation of DKI scalar measures and to improve the sensitivity to detect PD-related imaging features.
With CNN method, we found greater MD values in several brain regions in PD patients compared with HCs, especially in the left SN, which is consistent with previous reports using regions of interest (ROI) analysis 10, 11, 14, 30–32. PD is characterized by progressive death of dopaminergic neurons in the SN, followed by the loss of dopaminergic projections from the SN to striatum, resulting in a series of motor and non-motor symptoms 33. According to the mathematical concept of tensor, the three-dimensional shape of diffusion elliptical structure depends on three eigenvalues (λ1, λ2, λ3) of orthogonal principal axes without directions. The MD value is the average of the three eigenvalues. The impaired axons and neurons and loss of myelin integrity in PD patients result in the decrease of restriction of water molecules displacement, which induces increased MD values 32, 34. Regional increased MD values in the left SN and bilateral caudate estimated by CNN method are consistent with the pathological lesions in PD patients. In contrast, we did not find increased MD in the SN in PD patients compared to HCs by applying model fitting method. This finding indicates that CNN method can better reveal the pathological features of PD than model fitting method.
We did not observe modulation of FA and MK values in the SN in PD patients, which is in line with a previous report 14. In contrast, some previous studies based on ROI analysis showed decreased or increased FA and/or increased MK in the SN in PD patients 6, 31. We speculate that, first, different analysis methods may responsible for those controversial results. Whole-brain unpaired t-test, moving beyond the hypothesis-driven ROI analysis, focused the statistical information on each voxel accompanied with increased partial volume effects and false-positive risk, especially within the pathological brain tissues. Second, we suppose that these controversial findings may due to the heterogeneity of patients being recruited and variations imaging quality 3, 8. In addition, it has been reported that iron deposition could increase FA values and decrease MD values in the white and gray matters 35. Numerous reports have demonstrated that there was iron accumulation in the SN 36–38. Thus, different levels of iron deposition in the SN may also be a reason contributing to these inconsistent findings.
We found negative correlation between H&Y scales and CNN-estimated FA values and MK values in the SN, as well as positive correlation between H&Y scales and CNN-estimated FA values in the putamen. These results indicate that FA and MK in the SN decreases, while FA in the putamen increases with the progression of the disease. As most of our patients were in the early stages (fifty-five of our patients were at H&Y stages 1 and 2), it is possible we could detect decreased FA in the SN if more advanced patients were enrolled. We did not find any significant correlation between DKI scalar measures and clinical assessments in PD patients using the model fitting method, which further proves that using CNN method to estimate DKI measures can improve the ability to explore PD-related neural modulations than using model fitting method.
For the TBSS analysis, increased FA values were observed in the brain white matter such as anterior thalamic radiation, inferior longitudinal fasciculus, superior longitudinal fasciculus, corticospinal tract, and inferior fronto-occipital fasciculus with both methods and agreed with previous studies 13, 39–41. It has been shown that increased FA in these white matters correlated with better olfactory function and lower motor severity 42. Thus, these increased diffusional properties of white matter might be a reflection of microstructural compensation 42.
We observed greater MK values in white matter in PD patients, which is inconsistent with previous reports. Previous studies found either no significant difference in MK values 43, 44, or decreased MK values of anterior cingulum, inferior fronto-occipital fasciculus and uncinate fasciculus in PD patients 7, 45. We suppose that the heterogeneity of patients being recruited and differences in protocol of diffusion image acquisition and image processing may contribute to these inconsistent findings. In addition, we found increased KFA in white matter, which has not been reported previously. KFA values, resembling the FA definition, quantify the degree of anisotropy of the non-Gaussian diffusion. In the current study, the increased KFA and FA values are located in the same white matter fibers. So far, only a small number of studies have paid attention to the kurtosis changes of white matter in PD patients, it is necessary to perform large cohort studies to elucidate the microstructural changes in white matter in PD patients.
In conclusion, the CNN method has the potential to sensitively detect the nigral pathology and improve the robustness and performance of DKI images with few of DWIs and then to differentiate PD patents from HCs. In addition, compared with the model fitting method, CNN method can better find the relationship between DKI parameter measures and clinical assessments susceptibility. These findings approve that CNN can help to explore PD-associated imaging features.