Weakly supervised deep feature learning to predict amyloidosis from structural MR brain images

Presenting Author Senior Author
Name: Cyrus Manuel Name: Cyrus Manuel
Email: Email:
Presenting Author’s RIG/SRG:  
Presenting Author's Lab Location: VAMC   

Abstract Information
Imaging Modality: MR
Disease Application: Neurodegenerative Diseases
Complete author list: Cyrus Manuel; Duygu Tosun-Turgut
Abstract highlights: A deep learning framework on graphs derived from structural MR data was used to predict amyloid PET (AV45-PET) positivity with baseline clinical information. Hidden layers describe pertinent brain regions used for prediction and offers insight into key atrophy regions involved with Alzheimer’s disease staging.
Many novel treatments are aimed to target A𝜷, one of the pathological hallmarks of AD, but are hampered by potential non responders due to lack of target A𝜷 pathology in their brains. Specifically, about ~25-40% of those clinically diagnosed with AD or MCI would not have significant A𝜷 pathology. In this study we used a deep learning framework to predict A𝜷 pathology positivity from baseline clinical assessments and structural MRI data routinely acquired from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Graph convolutional networks (GCN) were trained on graphs derived from MR imaging and their performances was assessed to see their predictive value based on ground-truth A𝜷 positivity estimates from AV45 PET scans. Hidden layers of the first convolutional layer were visualized to observed pertinent regions the model has learned to be important towards predicting amyloidosis.
Baseline MRI and AV45-PET along with AD-related factors (age, gender, education, genotype, baseline cognitive performance) of 771 participants (248 healthy controls (HC), 381 mild cognitive impairment (MCI), 101 AD) from the ADNI database were used for evaluation. Undirected graph model reconstructed from diffusion MRI served as inputs for the GCNs. Anatomical brain parcellations with atrophy estimates from structural MRI constitute the vertices; tractography based connectivity estimates defined the edges of the graph model. Atrophy data were converted to standard scores relative to HC. Separate GCNs models were trained for each clinical diagnostic group. The best performing model architecture were used to assess model performance. Predictive value was compared with models trained on atrophy data and models with atrophy and AD-related factors added at the dense layer. The model assessments were performed on 10 independent training (60%) and test sets (40%). The diagnostic value of the numerical models was assessed by the mean test accuracy, sensitivity, specificity, and predictive value. A t-test was performed to determine significance (p>0.05). Outputs of learned filters in the first convolutional layer were visualized and analyzed.
GCNs were able to learn from atrophy descriptors and network connectivity derived from MRI and predict A𝜷 positivity . Atrophy was a significant predictor of A𝜷 positivity in the AD model, but at a lesser degree in HC and MCI models. The inclusion of other AD-related factors showed: a significant improvement in test accuracy and sensitivity in the AD model; a significant test accuracy, sensitivity, specificity, NPV, and PPV in MCI models; and a significant NPV and PPV in HC models. In the MCI model, filters of the first hidden layer suggest greater contribution of atrophy in left superior parietal, right inferior temporal, right entorihnal, and left postcentral regions in this network model.
Baseline structural MRI represented as a graph can offer predictive value to weather or not a subject will test positive for an AV45-PET exam. Predictions are more accurate with the addition of well-established AD-related factors, however more features may be necessary to increase the predictive ability in HC and MCI subjects. This assessment offers a practical adjunct to deciding the next course of action for the patient.