[Paper Review] On the design of convolutional neural networks for automatic detection of Alzheimer's disease

Sheng Liu, Chhavi Yadav+2 more|arXiv (Cornell University)|Nov 9, 2019

Brain Tumor Detection and ClassificationNeuroscience18 citations

TL;DR

This paper proposes a 3D convolutional neural network (CNN) architecture for automatic detection of Alzheimer’s disease (AD) using structural T1-weighted MRI scans, leveraging instance normalization, small-kernel initial layers to avoid early downsampling, wide network widths, and age embedding. The model achieves a 14% improvement in test accuracy over prior methods on the ADNI dataset and generalizes well to an independent AIBL dataset.

ABSTRACT

Early detection is a crucial goal in the study of Alzheimer's Disease (AD). In this work, we describe several techniques to boost the performance of 3D deep convolutional neural networks (CNNs) trained to detect AD using structural brain MRI scans. Specifically, we provide evidence that (1) instance normalization outperforms batch normalization, (2) early spatial downsampling negatively affects performance, (3) widening the model brings consistent gains while increasing the depth does not, and (4) incorporating age information yields moderate improvement. Together, these insights yield an increment of approximately 14% in test accuracy over existing models when distinguishing between patients with AD, mild cognitive impairment, and controls in the ADNI dataset. Similar performance is achieved on an independent dataset.

Motivation & Objective

• To improve the accuracy of automatic Alzheimer’s disease detection from structural brain MRI scans using deep learning.
• To address the challenge of low classification performance in existing models for distinguishing between cognitively normal (CN), mild cognitive impairment (MCI), and AD patients.
• To investigate architectural design choices that enhance 3D CNN performance on 3-class MRI classification.
• To ensure generalization by validating the model on an independent dataset (AIBL) with no data leakage.
• To provide insights into model interpretability through saliency map analysis of learned features.

Proposed method

• The model uses instance normalization instead of batch normalization to stabilize training and improve performance.
• It employs small-sized convolutional kernels in the first layer to avoid early spatial downsampling, preserving spatial resolution.
• The architecture is widened with a large number of filters per layer, while maintaining relatively few depth layers, to enhance representational capacity.
• Patient age is incorporated as a learned embedding, inspired by techniques from natural language processing, to improve classification.
• Saliency maps are computed as the gradient magnitude of the predicted class score with respect to the input MRI volume to visualize learned features.
• The model is trained and evaluated on the ADNI dataset with careful data splitting to prevent data leakage, and validated on the independent AIBL dataset.

Experimental results

Research questions

• RQ1 Does instance normalization outperform batch normalization in 3D CNNs for AD classification on MRI data?
• RQ2 How does early spatial downsampling affect the performance of 3D CNNs in detecting AD from structural MRI?
• RQ3 What is the relative impact of increasing network width versus depth on classification accuracy for AD detection?
• RQ4 Can incorporating patient age as a continuous embedding improve model performance in distinguishing CN, MCI, and AD?
• RQ5 Does the proposed architecture generalize effectively to an independent, external MRI dataset?

Key findings

• Instance normalization outperforms batch normalization, leading to improved model stability and accuracy in 3D CNNs for AD detection.
• Early spatial downsampling significantly degrades performance, and avoiding it through small-kernel initial layers yields consistent gains.
• Widening the network (increasing filter count) leads to consistent performance improvements, while increasing depth provides no such benefit.
• Incorporating patient age as an embedded feature results in a moderate but measurable improvement in classification accuracy.
• The proposed model achieves a 14% increase in test accuracy over existing models when distinguishing between CN, MCI, and AD on the ADNI dataset.
• The model generalizes effectively to an independent AIBL dataset, achieving comparable performance (63.6% accuracy) without using age information, confirming robust feature learning.