Source-linked AI summary
Towards Alzheimer's Disease Classification through Transfer Learning
Marcia Hon, Naimul Khan
TL;DR
Deep-learning AD diagnosis from MRI is limited by the need for large annotated datasets and extensive optimization. This paper addresses those constraints by transferring ImageNet weights to VGG16 and Inception, retraining the classifier, and selecting MRI slices by entropy. On OASIS data, the method achieves comparable or better performance than deep-learning approaches despite using a much smaller training set.
Problem
Deep-learning AD diagnosis requires many annotated MRI images and careful network optimization, which are difficult constraints in medical imaging.
Method
The paper transfers ImageNet pre-trained weights to VGG16 and Inception, fine-tunes the models, and selects informative MRI slices using image entropy.
Results
The method achieves comparable or better AD-detection performance than current deep-learning methods using a training set almost 10 times smaller than the state of the art.
Takeaways & Limitations
Transfer learning with entropy-based slice selection can support accurate AD detection from MRI with relatively few training images.
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
Automated MRI-based AD diagnosis is important, but deep learning from scratch is constrained by data, computation, and optimization demands. The paper investigates transfer learning with VGG16 and Inception, combined with entropy-based training-data selection, for small MRI datasets.
- AD diagnosis from MRI is an important application of machine learning because automated analysis may support early diagnosis and sometimes predict AD better than clinicians.
- Deep learning methods have surpassed earlier statistical approaches such as SVM and sparse autoencoders for automated AD detection.
- Training deep architectures from scratch requires extensive annotated data, substantial computation, and careful parameter tuning.
- Transfer learning reuses CNNs trained on large natural-image datasets and fine-tunes final fully-connected layers with smaller medical-imaging datasets.
- The paper adapts VGG16 and Inception to AD diagnosis, showing that ImageNet pre-trained weights generalize well when MRI training data are selected intelligently.
- Entropy-based selection chooses informative MRI slices, enabling comparable or better performance than training from scratch with minimal parameter optimization.
II. RELATED WORKS
Prior AD imaging work includes classical machine learning and increasingly successful deep learning methods. Transfer learning has also been applied across medical-imaging tasks, often outperforming training from scratch.
- Classical AD diagnosis methods use structural MRI features with classifiers such as SVM and feed-forward neural networks.
- Recent deep learning approaches include autoencoder features, convolutional layers, 3D convolution, stacked autoencoders, softmax classification, and CNN architectures such as LeNet and Inception.
- Medical-imaging studies have explored transfer learning by fine-tuning CNNs rather than training them from scratch.
- Fine-tuned CNNs have been used for ultrasound-plane localization, interstitial-lung-disease classification, and cardiac-imaging retrieval of missing or noisy views.
A. Convolutional Neural Networks and Transfer Learning
CNNs learn local image features through convolution and pooling, while transfer learning initializes networks with pre-trained weights and retrains only the final classifier. The paper uses VGG16 and Inception V4 as candidate architectures.
- CNN convolutional layers extract local features, while max pooling reduces feature-map size by selecting maximum local responses.
- Training from scratch randomly initializes weights and updates them through back-propagation to minimize a cost function.
- With small datasets, scratch training may lead to local minima and overfitting or underfitting; transfer learning instead initializes layers with pre-trained weights.
- Transfer learning retrains only the last fully-connected layer, providing pre-trained weights and established architectures for the target problem.
- VGG16 is a 16-layer architecture using small 3x3 filters, whereas Inception V4 uses global average pooling and softmax classification to reduce parameters and overfitting.
B. Most informative training data selection
Because small MRI training sets may not capture sufficient structural variation when sampled randomly, the paper selects slices by image entropy. Higher entropy is treated as greater variation and informativeness.
- The method calculates each MRI slice’s image entropy and selects the most informative slices instead of sampling training images randomly.
- Image entropy is computed from a slice’s histogram and measures variation within that slice.
- Sorting slices by entropy in descending order identifies the highest-entropy images as the training samples intended to provide robustness.
IV. EXPERIMENTAL RESULTS
The experiments evaluate VGG16 and Inception V4 CNNs for distinguishing Alzheimer’s disease from healthy controls using MRI scan data.
- IV. EXPERIMENTAL RESULTS: VGG16 and Inception V4 were evaluated for differentiating AD patients from Healthy Controls through MRI analysis.The deep-learning methods were implemented with Keras and TensorFlow, while training-data selection was implemented in MATLAB.
A. Dataset
The study uses cross-sectional OASIS MRI data, with example images representing Alzheimer’s disease and healthy-control groups.
- A. Dataset: The OASIS dataset provides MRI images from AD and HC groups, illustrated with representative images in Figure 1.The study uses structural MRI data from the Open Access Series of Imaging Studies.
- A. Dataset: The experiment randomly selected 200 subjects, comprising 100 AD subjects and 100 HC subjects, from 416 available subjects.The available subjects ranged in age from 18 to 96.
- A. Dataset: Cross-sectional rather than longitudinal OASIS data were used because the target was to differentiate AD from HC using images.The dataset contains both cross-sectional and longitudinal data.
B. Accuracy results
Accuracy was measured with 5-fold cross-validation across VGG16 and Inception V4 settings, including a VGG16-from-scratch baseline. The authors note that recommended hyperparameters performed well, while broader search was deferred because of computational cost.
- B. Accuracy results: 5-fold cross-validation used an 80%-20% training-testing split to obtain the reported results.This evaluation procedure was applied to the tested models.
- B. Accuracy results: VGG16 was trained both from scratch and with ImageNet-pretrained weights, while Inception V4 used pretrained weights for transfer learning.VGG16 used 100 epochs and batch size 40; Inception V4 used 100 epochs and batch size 8.
- B. Accuracy results: VGG16 used RMSProp, whereas Inception V4 used stochastic gradient descent with learning rate 0.0001.The RMSProp setting used an adaptive learning rate.
- B. Accuracy results: Grid search could potentially improve results, but it was not conducted because hyperparameter search was computationally expensive.The authors plan to explore it with cloud computing or additional hardware resources.
- B. Accuracy results: Transfer learning substantially improved accuracy over VGG16 trained from scratch, and Inception V4 with transfer learning achieved high accuracy.The authors attribute poor from-scratch performance to the small training size and possible overfitting or underfitting.
C. Comparison with other methods
The paper compares accuracy and training size with five recent methods while cautioning that datasets and experimental setups differ. Its approach achieves accuracy close to the best Inception-based comparator with a much smaller training set.
- C. Comparison with other methods: The comparison covers five recent methods, four of them deep-learning-based, using reported sample sizes and validation splits to estimate training size.The paper compares methods using structural MRI images, despite differences in datasets and experimental setups.
- C. Comparison with other methods: The proposed method outperforms every compared method except in accuracy.The comparison is presented in Table II.
- C. Comparison with other methods: 5,120 vs 46,751 training images: the proposed Inception-based method achieves accuracy very close to with a training set almost 10 times smaller.Both methods use the Inception architecture.
- C. Comparison with other methods: Transfer learning combined with entropy-based selection is reported as producing strong performance with a comparatively smaller training dataset.The authors state that the smaller dataset reduces dependence on extensively annotated data and improves training computational time.
V. CONCLUSION
The paper proposes transfer learning with VGG16 and Inception V4, using ImageNet pre-trained weights and entropy-based selection to detect AD from structural MRI with limited training data. On OASIS, the method achieves performance comparable to the state-of-the-art despite a much smaller training set.
- VGG16 and Inception V4 are fine-tuned from ImageNet pre-trained weights to detect AD from structural MRI images.The method also uses entropy-based selection to retain the most informative training images.
- The method provides performance comparable to the state-of-the-art despite having a training set many times smaller.
- Future work will evaluate the models on other datasets and further optimize hyperparameters using grid search.
- The models, dataset, and code are made available through a repository for reproducible research.