PaperlayerView exact revision on arXiv

Source-linked AI summary

3D CNN-based classification using sMRI and MD-DTI images for Alzheimer disease studies

Alexander Khvostikov, Karim Aderghal, Jenny Benois-Pineau, Andrey Krylov, Gwenaelle Catheline

arXiv:1801.05968v1cs.CV

TL;DR

The paper addresses imaging-based classification of Alzheimer’s disease and mild cognitive impairment, where CNNs face limited training data and low-resolution ROI inputs. It applies an adapted 3D CNN to hippocampal sMRI and MD-DTI, reporting accuracies of 0.967 for AD-NC, 0.8 for AD-MCI, and 0.658 for MCI-NC.

  • Problem

    For AD classification, deep neural networks face limited training data and low-resolution inputs when using ROIs.

    Source: §2.2 · p. 3

  • Method

    An adapted CNN classifies 3D hippocampal ROIs using aligned sMRI and MD-DTI, with class balancing during data augmentation.

    Source: §1 · p. 2, §3.1 · p. 6, §3.2 · p. 6

  • Results

    0.967 for AD-NC, 0.8 for AD-MCI, and 0.658 for MCI-NC were the reported classification accuracies; the authors also report fused 3D-network accuracy above 0.9.

    Source: §4 · p. 12, §5 · p. 12

  • Takeaways & Limitations

    The authors suggest CNN-based classification may be used in real-world CAD systems for large-cohort screening, while noting their study focused on the hippocampus.

    Source: §5 · p. 12

    AD classification with neural networks is constrained by limited training data and low-resolution ROI inputs.

    Source: §2.2 · p. 3

Abstract

from arXiv · show

Computer-aided early diagnosis of Alzheimers Disease (AD) and its prodromal form, Mild Cognitive Impairment (MCI), has been the subject of extensive research in recent years. Some recent studies have shown promising results in the AD and MCI determination using structural and functional Magnetic Resonance Imaging (sMRI, fMRI), Positron Emission Tomography (PET) and Diffusion Tensor Imaging (DTI) modalities. Furthermore, fusion of imaging modalities in a supervised machine learning framework has shown promising direction of research. In this paper we first review major trends in automatic classification methods such as feature extraction based methods as well as deep learning approaches in medical image analysis applied to the field of Alzheimer's Disease diagnostics. Then we propose our own algorithm for Alzheimer's Disease diagnostics based on a convolutional neural network and sMRI and DTI modalities fusion on hippocampal ROI using data from the Alzheimers Disease Neuroimaging Initiative (ADNI) database (http://adni.loni.usc.edu). Comparison with a single modality approach shows promising results. We also propose our own method of data augmentation for balancing classes of different size and analyze the impact of the ROI size on the classification results as well.

1. Introduction

AD-related brain degeneration can appear across several imaging modalities, motivating computer-aided diagnosis research. The paper reviews classification approaches and proposes a 3D CNN using hippocampal sMRI and DTI data from ADNI.

1. Introduction

  • sMRI, fMRI, PET, and DTI images are identified as sources for early detection, with hippocampal changes described as AD biomarkers.The introduction also notes that AD-related brain-cell degeneration produces structural changes visible in imaging.Source: §1 · p. 1
  • Earlier volumetric methods compared brain anatomy across subjects, while feature-based methods drew on local image features and CNNs were gradually replacing feature extraction approaches.The review notes that voxel-based morphometry compares local tissue concentrations with normal-control brains.Source: §1 · pp. 1–2
  • The paper surveys classification trends and proposes a supervised 3D CNN for hippocampal ROIs that fuses sMRI and DTI from an ADNI subset.The paper is organized to review prior methods before presenting its method, results, and discussion.Source: §1 · p. 2

2. Review of the existing classification methods in the problem of AD detection

Feature-based AD detection uses prior knowledge about ROI biomarkers and the characteristics of imaging signals.

2. Review of the existing classification methods in the problem of AD detection

  • Feature-based AD detection uses domain knowledge of ROI biomarkers and sMRI/DTI signal properties, which conventional SIFT and SURF descriptors do not sufficiently capture.The passage characterizes these signals as blurry.Source: §2 · p. 2

2.1. Feature-based classification

The review covers feature-based classification across sMRI, DTI, and multimodal fusion. The cited studies use varied descriptors and classifiers, with reported results differing across tasks and datasets.

2.1. Feature-based classification

  • Feature-based classification is reviewed for sMRI, DTI, and fusion with other modalities.The review frames these approaches as alternatives within image-based AD classification.Source: §2.1 · p. 2

2.1.1. sMRI

  • The sMRI study used Gauss-Laguerre Harmonic Functions and a bag-of-visual-words representation, reporting accuracies of 0.838 for AD/NC, 0.695 for NC/MCI, and 0.621 for AD/MCI.Its features were classified with an RBF-kernel SVM and 10-fold cross-validation on an ADNI subset.Source: §2.1.1 · p. 2

2.1.2. DTI

  • The DTI study represented connectivity among 41 Brodmann areas and reported accuracies of 0.8 for AD/NC, 0.833 for AD/MCI, and 0.7 for MCI/NC.It used FA maps and an ensemble of classifiers on a custom database.Source: §2.1.2 · p. 2
  • The DTI study reported accuracies of 0.986 for AD/MCI and 0.977 for AD/NC using FA and MO features with an SVM.The method used Relief feature selection and 10-fold cross-validation on 50 patients.Source: §2.1.2 · p. 2

2.1.3. Data fusion

  • The sMRI–PET fusion study used canonical correlation analysis to derive projected features, then reported accuracies of 0.969 for AD/NC and 0.866 for MCI/NC.The resulting projections were combined into a patient representation and classified with an SVM.Source: §2.1.3 · pp. 2–3
  • A separate fusion approach combined sMRI and MD-DTI visual features with CSF features using multiple-kernel learning, reporting accuracies of 0.902 for AD/NC, 0.794 for MCI/NC, and 0.766 for AD/MCI.The feature vectors were classified with an SVM-based multiple-kernel-learning approach.Source: §2.1.3 · p. 3

2.2. Classification with neural networks

The section reviews neural-network approaches to Alzheimer’s classification, including autoencoder pretraining, transfer learning, 2D and 3D CNNs, and other architectures. It highlights the constraints of small training datasets and low-resolution ROI inputs, alongside strategies explored to address them.

2.2. Classification with neural networks

  • Small training datasets and low-resolution ROI inputs are key constraints for deep networks in AD classification; proposed responses include shallow networks, transfer learning, and layer pretraining.The review also discusses domain-specific augmentation and using multiple ROIs or patches as ways to enlarge the data or input.Source: §2.2 · p. 3

2.2.1. Autoencoders

  • Stacked 3D convolutional autoencoders, sparse autoencoders, and sMRI/PET fusion are among the reviewed strategies, with reported classification accuracies varying by task and dataset.Reported examples include 0.993 for AD/NC with stacked 3D autoencoders, and 0.91 for AD/NC with sparse autoencoders and sMRI/PET fusion.Source: §2.2.1 · pp. 3–4

2.2.2. Transfer Learning

  • Transfer learning with ImageNet-pretrained AlexNet required converting 3D sMRI and PET data into resized 2D, three-channel inputs; another reviewed approach combined transfer learning with Tucker tensor decomposition for MCI classification.The AlexNet approach reported accuracies of 0.665 for AD/NC and 0.488 for AD/MCI/NC; the Tucker-based approach reported 0.906 on a 629-subject ADNI subset.Source: §2.2.2 · p. 4

2.2.3. 2D convolutional neural networks

  • Reviewed 2D CNN methods classify slices or hippocampal projections separately, with reported accuracy depending on the task and aggregation approach.One method combined three projection-specific networks at a fully connected layer and used majority voting; another selected 20 central slices and reported 0.983 for AD/NC, 0.939 for AD/MCI, and 0.917 for MCI/NC.Source: §2.2.3 · pp. 4–5

2.2.4. Other networks

  • A deep polynomial network used sMRI and PET ROI intensities and reported 0.971 accuracy for AD/NC, 0.872 for MCI/NC, and 0.789 for distinguishing converting from non-converting MCI.Its inputs were average intensities from 93 selected ROIs.Source: §2.2.4 · p. 5
  • Plain and residual 3D CNNs had nearly equal performance in one sMRI comparison, with 0.79–0.8 accuracy for AD/NC on an ADNI subset.That study examined four binary classification tasks involving AD, late MCI, early MCI, and normal controls.Source: §2.2.4 · p. 5
  • Other reviewed designs include residual-block ordinal ranking and spectral CNNs that represent brain ROIs as a graph, with performance reported across AD/NC, AD/MCI, and MCI/NC tasks.The ordinal-ranking method used hippocampal ROI data; the spectral CNN formed edges using Pearson correlation from grey matter.Source: §2.2.4 · p. 5
  • A sparse-regression/CNN approach converted multiple sparse representations of 93 sMRI ROIs into a matrix for CNN classification, reporting 0.903 accuracy for AD/NC and 0.742 for MCI/NC.The representations were varied using a sparse-control parameter before being fed to a CNN with two convolutional and two fully connected layers.Source: §2.2.4 · p. 5

3. Proposed method of classification

The method uses ADNI sMRI and MD-DTI scans to classify hippocampal regions with a 3D CNN, balancing classes through augmentation and exploring ROI size and network configurations.

3.1. Data selection

  • The 214 selected subjects comprise 48 AD, 108 MCI, and 58 NC patients, each with T1-weighted sMRI and DTI scans.The dataset size is limited by DTI availability, and the study focuses on hippocampal regions and surrounding tissue.Source: §3.1 · pp. 5–6
  • After DTI correction, skull stripping, and tensor fitting, MD maps are aligned to sMRI and both modalities are normalized to 121 × 145 × 121 voxels.Only MD maps are used; the aligned scans share the same resolution.Source: §3.1 · p. 6
  • Atlas AAL selects left and right hippocampal ROIs of 28 × 28 × 28 × 28 voxels, with larger centered ROIs tested to assess image-data volume.The extended ROI sizes are 38 × 38 × 38 × 38, 42 × 42 × 42 × 42, and 48 × 48 × 48 × 48 voxels.Source: §3.1 · p. 6

3.2. Data augmentation

  • Class balancing is performed through augmentation because MCI patients outnumber AD patients; each class is expanded to nmax · k samples using shifted and blurred images.New images use random shifts of up to 2 pixels in each dimension and Gaussian blur with σ up to 1.2.Source: §3.2 · pp. 6–7
  • Training uses k = 10 to produce 960 images per class, while test results are compared across original, shift-augmented, and shift-plus-blur sets.The three test sets contain 12, 120, and 120 samples, respectively.Source: §3.2 · p. 7

3.3. Network architecture

  • Separate 3D CNN pipelines process each ROI and modality, then concatenate flattened outputs for fully connected, dropout, and softmax layers.The pipeline uses convolution, batch normalization, rectifier linear unit, and pooling blocks; Figure 6 depicts the architecture and Figure 5 the convolutional block.Source: Figure 6 · p. 7, §3.3 · p. 7, Figure 5 · p. 7
  • Shallower C1/C2 networks are paired with 28- and 38-voxel ROIs, while deeper C3/C4 networks are used for 42- and 48-voxel ROIs.The architectures vary in convolutional depth and filter count.Source: §3.3 · p. 7, Table 4 · p. 8
  • To limit overfitting, training repeatedly uses a random 90% of training data for weight learning and the remaining 10% for validation.The train-validation split is repeated after a fixed number of iterations.Source: §3.3 · p. 7

4. Results

The experiments compare three binary classification tasks across imaging inputs, ROI sizes, and network configurations. The reported results favor fusion for AD-NC and AD-MCI, while larger ROIs, test-set augmentation, and balanced test classes shape how performance is interpreted.

4. Results

  • Experiments compared AD-NC, AD-MCI, and MCI-NC across modality inputs, ROI sizes, and network configurations, reporting accuracy, sensitivity, and specificity.Source: §4 · p. 8
  • Because test classes had equal patient counts, reported accuracy equaled balanced accuracy; metrics used top-mean values and 95% Wilson score intervals.Top-mean accuracy was the maximum mean across sequential 100-iteration intervals; interval width varied with sample count.Source: §4 · p. 8
  • The third input type—separate left and right hippocampal MD-DTI and sMRI ROIs—performed best for AD-NC and AD-MCI, while fusion was slightly inferior for MCI-NC.The authors attribute the MCI-NC result to less informative DTI adding noise to fully connected layer weights with limited data.Source: §4 · pp. 8, 12
  • Larger 38×38×38 and 48×48×48 ROIs with deeper networks improved classification; 48×48×48 with a six-layer fused network achieved 0.967 accuracy.A 38×38×38 ROI with the five-layer C2 architecture also performed well across all three tasks.Source: §4 · p. 12
  • More test-set augmentation produced smoother accuracy curves but slightly worsened classification results in all cases.Source: Figure 7 · p. 8, Table 5 · p. 9, Table 6 · p. 10
  • Accuracy reached 0.967 for AD-NC, 0.8 for AD-MCI, and 0.658 for MCI-NC classification.Source: §4 · p. 12

5. Discussion and Conclusion

The authors caution that reviewed algorithms cannot be directly compared because they use different datasets, sample sizes, and classification tasks. Their proposed 3D fused CNN achieved accuracy above 0.9, but this study focused on the hippocampal ROI alone.

5. Discussion and Conclusion

  • Direct comparison of reviewed algorithms is impossible because studies used different datasets, sample quantities, and classification problems.Source: §5 · p. 12
  • The authors report that 3D-volume deep neural networks with fused modalities achieved accuracy above 0.9, which they view as promising for large-cohort screening CAD systems.Source: §5 · p. 12
  • This study focused on the hippocampal ROI; the authors identify adding other AD-affected ROIs as a direction for further work.Source: §5 · p. 12
Loading 1801.05968v1…