PROJECT TITLE :

Multi-Task Weakly-Supervised Attention Network for Dementia Status Estimation With Structural MRI

ABSTRACT:

The accurate prediction of clinical scores (of neuropsychological tests) based on noninvasive structural magnetic resonance imaging (MRI) is helpful in understanding the pathological stage of dementia (such as Alzheimer's disease (AD)) and forecasting the progression of the disease. Pre-selection of dementia-sensitive brain regions for MRI feature extraction and model construction is common in existing machine learning and deep learning approaches. This can potentially lead to undesirable heterogeneity between different stages and degraded prediction performance. In addition, these methods typically rely on prior anatomical knowledge (for example, a brain atlas) and time-consuming nonlinear registration in order to preselect brain locations. As a result, they ignore the individual-specific structural changes that occur during the progression of dementia because all subjects share the same preselected brain regions. In this article, we propose a multi-task weakly-supervised attention network (MWAN) for the joint regression of multiple clinical scores based on baseline MRI scans. This network would be used to analyze the data. MWAN is comprised of three sequential components: 1) a backbone fully convolutional network for extracting MRI features; 2) a weakly supervised dementia attention block for automatically identifying subject-specific discriminative brain locations; and 3) an attention-aware multitask regression block for jointly predicting multiple clinical scores. Each of these components is described in greater detail below. The dementia-aware holistic feature learning and multitask regression model construction are integrated into a unified framework within the proposed MWAN, which is an end-to-end and fully trainable deep learning model. In order to estimate clinical scores on the mini-mental state examination (MMSE), clinical dementia rating sum of boxes (CDRSB), and the AD assessment scale cognitive subscale, our MWAN method was tested on two public AD data sets (ADAS-Cog). The quantitative experimental results demonstrate that our method generates superior regression performance when compared with methods that are considered to be state-of-the-art. Importantly, qualitative findings suggest that the dementia-sensitive brain locations automatically identified by our MWAN method do a good job of preserving the individual characteristics of each person and have biological significance.