[Paper Review] Towards Alzheimer's Disease Classification through Transfer Learning

Marcia Hon, Naimul Khan|arXiv (Cornell University)|Nov 29, 2017

Brain Tumor Detection and Classification21 citations

TL;DR

This paper proposes a transfer learning approach using pre-trained VGG16 and Inception V4 models on MRI data for Alzheimer’s disease classification. By selecting the most informative training slices using image entropy, the method achieves 96.25% accuracy with only 5,120 training images—demonstrating state-of-the-art performance using a training set nearly 10 times smaller than prior deep learning methods.

ABSTRACT

Detection of Alzheimer's Disease (AD) from neuroimaging data such as MRI through machine learning have been a subject of intense research in recent years. Recent success of deep learning in computer vision have progressed such research further. However, common limitations with such algorithms are reliance on a large number of training images, and requirement of careful optimization of the architecture of deep networks. In this paper, we attempt solving these issues with transfer learning, where state-of-the-art architectures such as VGG and Inception are initialized with pre-trained weights from large benchmark datasets consisting of natural images, and the fully-connected layer is re-trained with only a small number of MRI images. We employ image entropy to select the most informative slices for training. Through experimentation on the OASIS MRI dataset, we show that with training size almost 10 times smaller than the state-of-the-art, we reach comparable or even better performance than current deep-learning based methods.

Motivation & Objective

• To address the challenge of limited annotated MRI data in Alzheimer’s disease classification.
• To reduce dependency on large, expensive, and hard-to-acquire medical training datasets.
• To improve deep learning performance on small medical imaging datasets through transfer learning and intelligent data selection.
• To evaluate whether pre-trained models from natural image datasets generalize effectively to brain MRI for AD detection.
• To demonstrate that high accuracy can be achieved with minimal hyperparameter tuning and reduced computational cost.

Proposed method

• Fine-tuning pre-trained VGG16 and Inception V4 models using ImageNet weights on a small MRI dataset.
• Selecting the most informative MRI slices using image entropy to maximize information content in the training set.
• Training only the fully-connected layers of the CNNs while freezing the convolutional feature extractors.
• Performing 5-fold cross-validation to evaluate model robustness and generalization.
• Using the OASIS dataset with 6,400 MRI slices from 200 subjects, preprocessed and normalized.
• Employing standard deep learning training protocols with Adam optimizer and categorical cross-entropy loss.

Experimental results

Research questions

• RQ1 Can transfer learning with pre-trained ImageNet models achieve high accuracy in Alzheimer’s disease classification using a small MRI dataset?
• RQ2 Does entropy-based selection of MRI slices improve model performance compared to random sampling?
• RQ3 How does the performance of transfer learning compare to training from scratch on the same small dataset?
• RQ4 Can a smaller training set with intelligent selection outperform larger datasets used in prior deep learning methods?
• RQ5 To what extent does the choice of architecture (VGG16 vs Inception V4) impact performance under transfer learning with limited data?

Key findings

• The Inception V4 model fine-tuned with transfer learning achieved 96.25% accuracy with a training set of only 5,120 images.
• Transfer learning significantly outperformed training VGG16 from scratch, which achieved only 74.12% accuracy.
• The proposed method achieved higher accuracy than four out of five state-of-the-art deep learning methods, despite using a training set nearly 10 times smaller than the largest baseline.
• The entropy-based data selection strategy enhanced model performance by prioritizing structurally diverse and information-rich MRI slices.
• The method achieved competitive performance with minimal hyperparameter tuning and reduced computational cost.
• The results demonstrate that transfer learning combined with intelligent data selection can drastically reduce reliance on large annotated medical datasets.