Prediction of cognitive function based on structural MRI images using a 3D convolutional neural net among cognitively normal older adults

Abstract

<h3>Introduction</h3> It is projected that 30 years from now, more than 12 million US adults may be affected by neurodegenerative diseases such as Alzheimer's Disease. With no cure on the market and the looming societal pressure to care for the burgeoning numbers of elderly patients, the goal of modern medicine should be to identify patients at risk for dementia and intervene early on. Machine learning algorithms have been successful at classifying which patients have Alzheimer's Disease and which patients do not based off an MRI and/or PET image. Although neurodegenerative diseases are quite prevalent in elderly individuals, no machine learning algorithm has been developed to predict which healthy adults are at heightened risk for developing neurodegenerative diseases. Here, we take the first step towards this goal by predicting an individual's current cognitive function based off of a singular MRI. We drew inspiration from 3D convolutional neural nets (CNNs) to predict a cognitive test score at present based on a raw MRI image among cognitively normal older adults. <h3>Methods</h3> We analyzed data from 159 community dwelling cognitively normal individuals (mean age = 76.3 yo ± 6.5, 34% male). A neural net was designed to include four convolutional layers that used 3x3x3 filters. Each layer generated eight, 16, 32, and 64 feature maps respectively and each layer was normalized and followed by a max pooling layer. The convolutional layers were followed by two fully connected layers with 128 and 64 neurons respectively and the final output involved one output neuron that used a rectilinear activation function. Mean squared error loss was used to adjust the weights of the neural net, and the learning rate was set to 0.001. The inputs to the CNN were 159 skull stripped MPRAGE MR-T1 weighted images registered to MNI space and randomly rotated and translated (probability = 0.3). 20% of the data was held out for validation. The outcome variables were composite cognitive score and subdomain scores (i.e., memory-retrieval, memory-learning, language, visuospatial, and executive attention) which were derived from 21 different neuropsychological tests grouped according to the literature to reflect cognitive function across five different cognitive domains. The cognitive scores computed amongst all five subdomains were normally distributed. A cross-validated linear regression to predict cognitive function from age served as a comparative model since it is a clinically popular model. <h3>Results</h3> The 3D CNN demonstrated that when age was a statistically significant predictor of cognitive subdomain score or composite cognitive score (p < 0.001), the ratio of the mean squared error loss of a simple linear regression to the mean squared error loss from the CNN ranged from 1.10 to 1.47. This was true for the composite cognitive score as well as the executive attention and visuospatial subdomain scores. However, when age was not a statistically significant predictor of cognition, as was the case for memory-learning, memory-retrieval and language subdomain scores, the range of the ratio increased to 1.59 to 2.23. The most relevant areas of the brain that predicted cognition were cortical areas of the brain that are most sensitive to neurodegeneration with age including the cerebellum, frontal, and temporal lobes. <h3>Conclusions</h3> When age is not a statistically significant correlate of cognitive function, machine learning is at least 1.5 times better than a linear regression model at predicting cognitive function based on a single MRI. Furthermore, the areas of the brain that are most significant in predicting cognition in cognitively normal older adults provide both validation of our methodology as well as new insights. The 3D CNN gained the most amount of insight from the mid frontal lobe and inferior temporal lobe, brain regions known to be susceptible to neurodegeneration associated with aging. The 3D CNN, without any such foresight, replicated what we know to be already true. Additionally, the 3D CNN gained considerable insight from the cerebellum as well. Perhaps the significance was given to the frontal portion of the cerebellum as it could foreshadow future frontal cortex degeneration. Overall, the 3D CNN developed here was successful at predicting cognition off a singular MRI, outperforms current linear regression models, affirms the current state of knowledge regarding brain regions correlated with cognition, and provides novel insight into which brain regions are predictive of cognitive function. The model provides hope that it can be used as a tool to predict future cognitive decline in an elderly individual. The ability to predict future cognition would allow clinicians to identify those patients who are at considerable risk for abnormally steep changes in cognitive function and provide preemptive care to individuals more at risk for neurodegenerative diseases. <h3>Funding</h3> University of Pittsburgh School of Medicine Physician Scientist Training Program