Source-linked AI summary
Automatic Assessment of Alzheimer's Disease Diagnosis Based on Deep Learning Techniques
Alejandro Puente-Castro, Enrique Fernandez-Blanco, Alejandro Pazos, Cristian R. Munteanu
TL;DR
Early diagnosis of Alzheimer’s disease is important because treatment is palliative, and MRI reveals changes in brain morphology. The paper evaluates deep-learning diagnosis using sagittal MRI and transfer learning, reporting results comparable to horizontal-plane MRI work, including for early stages.
Problem
Early AD diagnosis matters because the disease has no cure beyond palliative treatments to slow its progression.
Source: §1 · p. 2Method
The study uses sagittal MRI from two reference datasets, extracting image features with ResNet and concatenating patient age and sex before training an SVM.
Source: §2.1 · pp. 3–4, §4 · p. 8Results
Takeaways & Limitations
The findings support investigating sagittal MRI for AD identification, and the model detected cognitively normal cases and early disease stages better in both datasets.
Source: §4 · p. 9
Abstract
from arXiv · showhide
Early detection is crucial to prevent the progression of Alzheimer's disease (AD). Thus, specialists can begin preventive treatment as soon as possible. They demand fast and precise assessment in the diagnosis of AD in the earliest and hardest to detect stages. The main objective of this work is to develop a system that automatically detects the presence of the disease in sagittal magnetic resonance images (MRI), which are not generally used. Sagittal MRIs from ADNI and OASIS data sets were employed. Experiments were conducted using Transfer Learning (TL) techniques in order to achieve more accurate results. There are two main conclusions to be drawn from this work: first, the damages related to AD and its stages can be distinguished in sagittal MRI and, second, the results obtained using DL models with sagittal MRIs are similar to the state-of-the-art, which uses the horizontal-plane MRI. Although sagittal-plane MRIs are not commonly used, this work proved that they were, at least, as effective as MRI from other planes at identifying AD in early stages. This could pave the way for further research. Finally, one should bear in mind that in certain fields, obtaining the examples for a data set can be very expensive. This study proved that DL models could be built in these fields, whereas TL is an essential tool for completing the task with fewer examples.
1 Introduction
Early diagnosis matters because Alzheimer’s disease has no cure beyond palliative treatments to slow progression, and MRI can reveal disease-related changes in brain morphology. Against prior work centered on horizontal MRI images or volumes, this study proposes exclusive sagittal imaging with fine-tuned general-purpose architectures.
1 Introduction
- MRI is used to visualize brain-morphology alterations associated with degeneration, while early detection matters because AD has no cure beyond palliative treatments to slow progression.In MRI, affected regions appear darker than healthy parts.Source: §1 · p. 2
- Prior image-based diagnostic work mainly used horizontal MRI images or volumes, whereas this study proposes using sagittal images exclusively with fine-tuned general-purpose architectures.The authors argue this could allow more exhaustive experiments at lower computational cost by using information from different vertical brain levels without processing many images.Source: §1 · p. 2
- The introduction notes that some prior studies used class-by-class binary comparisons, which the authors say can yield biased values because other classes’ noise is absent.Source: §1 · p. 3
- The study aims to assess AD diagnosis from sagittal MRI, establish that sagittal images can identify AD without another acquisition plane, and test transfer learning with fewer data.Source: §1 · p. 3
2 Materials and Methods
The study processes sagittal MRI from OASIS and ADNI with a ResNet-to-SVM workflow that incorporates patient age and sex. It evaluates the model using five confusion-matrix measures.
2.1 Workflow
- ResNet features, concatenated with patient age and sex, feed an RBF-kernel SVM trained and evaluated on separate training and testing data.The workflow extracts 100,352 values per MRI scan; transfer learning uses a ResNet trained on another task to reduce the weights and new data needed.Source: §2.1 · pp. 3–4, §2.3 · p. 4
2.2 Data Sets
- The study uses labeled MRI collections from OASIS and ADNI, two datasets commonly used in the field.Source: §2.2 · p. 4
2.2.1 OASIS
- OASIS-Cross-sectional supplies 436 sagittal images: 2 moderate AD, 28 mild dementia, 70 very mild dementia, and 316 cognitively normal cases.The set is strongly imbalanced toward cognitively normal cases; images come from 168 male and 268 female patients aged 18–98.Source: §2.2.1 · p. 4
2.2.2 ADNI
- ADNI contributes 1,743 sagittal images from the MP-RAGE REPEAT collection, split among 297 AD, 921 mild cognitive dementia, and 525 cognitively normal cases.The dataset is imbalanced in favor of AD; images are the central sagittal cut from volumes.Source: §2.2.2 · p. 4
2.3 Proposed Method
- The proposed model uses the first 47 layers of ImageNet-trained ResNet to extract MRI features, adds sex and age, then classifies with an SVM.ResNet is described as using residual blocks to address gradient problems in deep networks; transfer learning supports adapting previously trained models to this task.Source: §2.1 · p. 3, §2.3 · p. 4
2.4 Performance Measures
- The model is assessed with accuracy, precision, recall, specificity, and F1 score, chosen to support comparison with bioinformatics studies.Source: §2.4 · p. 4
3 Results
The results show that sagittal-MRI models diagnosed early-stage AD in both data sets, while performance on less-represented advanced stages was limited. Class weighting addressed imbalance, and the two data sets used different evaluation protocols.
3.1 Imbalance
- Class weights were set inversely to class representation, giving the least-represented class the greatest weight to counter imbalance and reduce overfitting to the majority class.For a dataset with 8 examples of class 1 and 2 of class 0, the respective weights were 0.625 and 2.5.Source: §3.1 · pp. 5–6
3.2 OASIS
- OASIS used Leave-One-Out evaluation with 1% of training data for validation, early stopping, and random-search hyperparameter optimization.The authors report that these searches helped prevent overfitting and memorization of the majority class.Source: §3.1 · p. 5, §3.2 · p. 6
- OASIS yielded 0% recall for its two advanced stages, with only 2 moderate-AD and 28 mild-dementia cases; adding sex and age slightly improved results for the first two classes.The model showed capacity to diagnose the earliest stage; data augmentation was used to address later stages at the cost of detecting fewer earliest-stage cases.Source: §3.2 · p. 6
- OASIS testing accuracy remained high as training examples increased from 10% to 70%, which the authors interpret as evidence of varied cases despite the dataset's small size.Figure 4 presents the accuracy learning curve for those subsets.Source: §3.2 · p. 7, Figure 4 · p. 7
3.3 ADNI
- ADNI used 50 repetitions of Hold-Out evaluation, splitting the data 80% for training and 20% for testing, with 1% of training data for validation.Training used early stopping and random-search hyperparameter optimization.Source: §3.3 · p. 7
- On ADNI, the AD stage was least correctly detected, with 30.62% recall, while results were better than OASIS because ADNI had less imbalance and more data.Considering demographic variables worsened results except for recall in the earliest stage, making healthy cases more likely to be identified as sick.Source: §3.3 · pp. 7–8
- ADNI's accuracy learning curve decreased for tests as training examples increased, which the authors suggest may reflect limited stage representation, noise, or overlapping stage cases.They note that these possibilities affect model learning.Source: §3.3 · p. 8, Figure 5 · p. 8
4 Discussion
The model identifies AD from sagittal MRI in both datasets, with results comparable to horizontal-plane work. It detects cognitively normal cases and early disease stages better, while class weighting addresses data imbalance.
4 Discussion
- The proposed DL model was evaluated on two reference datasets using class weights to address imbalance and established evaluation strategies and metrics.The authors report satisfactory results with sagittal images in both datasets.Source: §4 · p. 8
- Results from both datasets were comparable to horizontal-plane work despite data imbalance and OASIS’s small size, with class weighting cited as the reason for satisfactory performance.The authors attribute the model’s performance to empirically studying class weights during training.Source: §4 · p. 8
- The authors recommend limiting model depth and training to reduce memorization and overfitting, which they say can worsen results on new data.They also identify Early Stopping and weight adjustment as useful techniques.Source: §4 · pp. 8–9
- In both datasets, the model detects cognitively normal cases and early disease stages better, supporting its intended use in early diagnosis.The authors state that early routines and treatments can slow AD progression and help prevent later stages.Source: §4 · p. 9
- The authors report that sagittal MRI shows typical AD-related deformations, potentially allowing investigation of features in other regions to improve diagnostic accuracy.They frame this as a basis for further experimentation.Source: §4 · p. 9
5 Conclusion
The study applies transfer learning to sagittal MRI for AD detection, using a ResNet feature extractor and SVM classifier. Its results are satisfactory relative to horizontal-plane MRI work, especially for early stages, and the authors propose extensions using MRI volumes and additional planes.
5 Conclusion
- The study used transfer learning with a ResNet feature extractor and SVM classifier, testing the model on two reference datasets.The authors assessed goodness-of-fit using previously agreed evaluation strategies and metrics.Source: §5 · p. 9
- The model’s results were satisfactory compared with previous work using horizontal-plane MRI, especially for detecting early AD stages.The authors note that early stages are harder to detect because phenotypic manifestations are limited.Source: §5 · p. 9
- Transfer learning supports experiments with little data and faster experiment design, without the stated risk of unrealistic augmented cases or replicated labeling errors.The authors attribute the faster processing to pretrained models requiring fewer data.Source: §5 · p. 9
- Future work will combine sagittal MRI volumes and potentially frontal-plane information, and investigate which brain regions are affected earlier.The authors expect whole-brain sagittal information may improve predictions.Source: §5 · p. 9