Deep transfer learning of structural magnetic resonance imaging fused with blood parameters improves brain age prediction.

Machine learning has been applied to neuroimaging data for estimating brain age and capturing early cognitive impairment in neurodegenerative diseases. Blood parameters like neurofilament light chain are associated with aging. In order to improve brain age predictive accuracy, we constructed a model based on both brain structural magnetic resonance imaging (sMRI) and blood parameters. Healthy subjects (n = 93; 37 males; aged 50-85 years) were recruited. A deep learning network was firstly pretrained on a large set of MRI scans (n = 1,481; 659 males; aged 50-85 years) downloaded from multiple open-source datasets, to provide weights on our recruited dataset. Evaluating the network on the recruited dataset resulted in mean absolute error (MAE) of 4.91 years and a high correlation (r = .67, p <.001) against chronological age. The sMRI data were then combined with five blood biochemical indicators including GLU, TG, TC, ApoA1 and ApoB, and 9 dementia-associated biomarkers including ApoE genotype, HCY, NFL, TREM2, Aß40, Aß42, T-tau, TIMP1, and VLDLR to construct a bilinear fusion model, which achieved a more accurate prediction of brain age (MAE, 3.96 years; r = .76, p <.001). Notably, the fusion model achieved better improvement in the group of older subjects (70-85 years). Extracted attention maps of the network showed that amygdala, pallidum, and olfactory were effective for age estimation. Mediation analysis further showed that brain structural features and blood parameters provided independent and significant impact. The constructed age prediction model may have promising potential in evaluation of brain health based on MRI and blood parameters.

Feasibility of multi-parametric magnetic resonance imaging combined with machine learning in the assessment of necrosis of osteosarcoma after neoadjuvant chemotherapy: a preliminary study.

BACKGROUND: Response evaluation of neoadjuvant chemotherapy (NACT) in patients with osteosarcoma is significant for the termination of ineffective treatment, the development of postoperative chemotherapy regimens, and the prediction of prognosis. However, histological response and tumour necrosis rate can currently be evaluated only in resected specimens after NACT. A preoperatively accurate, noninvasive, and reproducible method of response assessment to NACT is required. In this study, the value of multi-parametric magnetic resonance imaging (MRI) combined with machine learning for assessment of tumour necrosis after NACT for osteosarcoma was investigated. METHODS: Twelve patients with primary osteosarcoma of limbs underwent NACT and received MRI examination before surgery. Postoperative tumour specimens were made corresponding to the transverse image of MRI. One hundred and two tissue samples were obtained and pathologically divided into tumour survival areas (non-cartilaginous and cartilaginous tumour viable areas) and tumour-nonviable areas (non-cartilaginous tumour necrosis areas, post-necrotic tumour collagen areas, and tumour necrotic cystic/haemorrhagic and secondary aneurismal bone cyst areas). The MRI parameters, including standardised apparent diffusion coefficient (ADC) values, signal intensity values of T2-weighted imaging (T2WI) and subtract-enhanced T1-weighted imaging (ST1WI) were used to train machine learning models based on the random forest algorithm. Three classification tasks of distinguishing tumour survival, non-cartilaginous tumour survival, and cartilaginous tumour survival from tumour nonviable were evaluated by five-fold cross-validation. RESULTS: For distinguishing non-cartilaginous tumour survival from tumour nonviable, the classifier constructed with ADC achieved an AUC of 0.93, while the classifier with multi-parametric MRI improved to 0.97 (P = 0.0933). For distinguishing tumour survival from tumour nonviable, the classifier with ADC achieved an AUC of 0.83, while the classifier with multi-parametric MRI improved to 0.90 (P < 0.05). For distinguishing cartilaginous tumour survival from tumour nonviable, the classifier with ADC achieved an AUC of 0.61, while the classifier with multi-parametric MRI parameters improved to 0.81(P < 0.05). CONCLUSIONS: The combination of multi-parametric MRI and machine learning significantly improved the discriminating ability of viable cartilaginous tumour components. Our study suggests that this method may provide an objective and accurate basis for NACT response evaluation in osteosarcoma.

Deep learning model to differentiate Crohn's disease from intestinal tuberculosis using histopathological whole slide images from intestinal specimens.

Crohn's disease (CD) and intestinal tuberculosis (ITB) share similar histopathological characteristics, and differential diagnosis can be a dilemma for pathologists. This study aimed to apply deep learning (DL) to analyze whole slide images (WSI) of surgical resection specimens to distinguish CD from ITB. Overall, 1973 WSI from 85 cases from 3 centers were obtained. The DL model was established in internal training and validated in external test cohort, evaluated by area under receiver operator characteristic curve (AUC). Diagnostic results of pathologists were compared with those of the DL model using DeLong's test. DL model had case level AUC of 0.886, 0.893 and slide level AUC of 0.954, 0.827 in training and test cohorts. Attention maps highlighted discriminative areas and top 10 features were extracted from CD and ITB. DL model's diagnostic efficiency (AUC = 0.886) was better than junior pathologists (*1 AUC = 0.701, P = 0.088; *2 AUC = 0.861, P = 0.788) and inferior to senior GI pathologists (*3 AUC = 0.910, P = 0.800; *4 AUC = 0.946, P = 0.507) in training cohort. In the test cohort, model (AUC = 0.893) outperformed senior non-GI pathologists (*5 AUC = 0.782, P = 0.327; *6 AUC = 0.821, P = 0.516). We developed a DL model for the classification of CD and ITB, improving pathological diagnosis accuracy effectively.

To explore the mechanism of tobacco addiction using structural and functional MRI: a preliminary study of the role of the cerebellum-striatum circuit.

Previous studies have found that the striatum and the cerebellum played important roles in nicotine dependence, respectively. In heavy smokers, however, the effect of resting-state functional connectivity of cerebellum-striatum circuits in nicotine dependence remained unknown. This study aimed to explore the role of the circuit between the striatum and the cerebellum in addiction in heavy smokers using structural and functional magnetic resonance imaging. The grey matter volume differences and the resting-state functional connectivity differences in cerebellum-striatum circuits were investigated between 23 heavy smokers and 23 healthy controls. The cigarette dependence in heavy smokers and healthy controls were evaluated by using Fagerström Test. Then, we applied mediation analysis to test whether the resting-state functional connectivity between the striatum and the cerebellum mediates the relationship between the striatum morphometry and the nicotine dependence in heavy smokers. Compared with healthy controls, the heavy smokers' grey matter volumes decreased significantly in the cerebrum (bilateral), and increased significantly in the caudate (bilateral). Seed-based resting-state functional connectivity analysis showed significantly higher resting-state functional connectivity among the bilateral caudate, the left cerebellum, and the right middle temporal gyrus in heavy smokers. The cerebellum-striatum resting-state functional connectivity fully mediated the relationship between the striatum morphometry and the nicotine dependence in heavy smokers. Heavy smokers showed abnormal interactions and functional connectivity between the striatum and the cerebellum, which were associated with the striatum morphometry and nicotine dependence. Such findings could provide new insights into the neural correlates of nicotine dependence in heavy smokers.

Characterizing the Structural Pattern of Heavy Smokers Using Multivoxel Pattern Analysis.

Background: Smoking addiction is a major public health issue which causes a series of chronic diseases and mortalities worldwide. We aimed to explore the most discriminative gray matter regions between heavy smokers and healthy controls with a data-driven multivoxel pattern analysis technique, and to explore the methodological differences between multivoxel pattern analysis and voxel-based morphometry. Methods: Traditional voxel-based morphometry has continuously contributed to finding smoking addiction-related regions on structural magnetic resonance imaging. However, voxel-based morphometry has its inherent limitations. In this study, a multivoxel pattern analysis using a searchlight algorithm and support vector machine was applied on structural magnetic resonance imaging to identify the spatial pattern of gray matter volume in heavy smokers. Results: Our proposed method yielded a voxel-wise accuracy of at least 81% for classifying heavy smokers from healthy controls. The identified regions were primarily located at the temporal cortex and prefrontal cortex, occipital cortex, thalamus (bilateral), insula (left), anterior and median cingulate gyri, and precuneus (left). Conclusions: Our results suggested that several regions, which were seldomly reported in voxel-based morphometry analysis, might be latently correlated with smoking addiction. Such findings might provide insights for understanding the mechanism of chronic smoking and the creation of effective cessation treatment. Multivoxel pattern analysis can be efficient in locating brain discriminative regions which were neglected by voxel-based morphometry.