Source-linked AI summary
Residual and Plain Convolutional Neural Networks for 3D Brain MRI Classification
Sergey Korolev, Amir Safiullin, Mikhail Belyaev, Yulia Dodonova
TL;DR
MRI classification studies often rely on multistep feature-extraction pipelines and face small neuroimaging datasets. This paper evaluates plain and residual 3D CNNs on ADNI scans, reporting comparable performance without handcrafted feature generation while noting that future work must address unpreprocessed images.
Problem
MRI classification commonly uses complex multistep feature-extraction pipelines, while neuroimaging datasets are small for supervised deep learning.
Method
The paper compares plain and residual 3D convolutional neural networks for classifying ADNI structural MRI scans across AD, EMCI, LMCI, and NC categories.
Results
The proposed models achieve relatively good and comparable performance to previous approaches without complex preprocessing or model stacking.
Takeaways & Limitations
The approach simplifies MRI classification by eliminating the need for handcrafted feature generation and supports direct application to 3D MRI images.
Takeaways & Limitations
Future work is needed to determine whether similar or better results can be achieved on images without alignment and skull-stripping preprocessing.
Abstract
from arXiv · showhide
In the recent years there have been a number of studies that applied deep learning algorithms to neuroimaging data. Pipelines used in those studies mostly require multiple processing steps for feature extraction, although modern advancements in deep learning for image classification can provide a powerful framework for automatic feature generation and more straightforward analysis. In this paper, we show how similar performance can be achieved skipping these feature extraction steps with the residual and plain 3D convolutional neural network architectures. We demonstrate the performance of the proposed approach for classification of Alzheimer's disease versus mild cognitive impairment and normal controls on the Alzheimer's Disease National Initiative (ADNI) dataset of 3D structural MRI brain scans.
1. INTRODUCTION
The paper addresses complex MRI classification pipelines and small neuroimaging datasets by proposing end-to-end plain and residual 3D CNNs for ADNI-based classification. It evaluates these models across Alzheimer’s disease, mild cognitive impairment, and normal-control scans, comparing them with prior approaches.
- Prior MRI-classification studies commonly use multistep pipelines for handcrafted feature generation and extraction before standard classifiers.
- Small neuroimaging datasets require architectures capable of learning classification features from limited supervised data.
- The paper proposes plain and residual 3D convolutional architectures that generate features automatically without intermediate handcrafted extraction.
- The models are evaluated on ADNI structural MRI scans for AD, EMCI, LMCI, and NC classification tasks.
- Prior work includes complex multimodal or feature-extraction systems, including DSA-3D-CNN, which reported binary ROC AUC above .96 but has weight-initialization time costs.
- The approach targets binary one-versus-one comparisons because previous ADNI studies primarily report binary classification results.
2. METHOD
The method compares a plain feedforward 3D CNN with a residual network, tuning their architectures and training settings for brain MRI classification. Evaluation uses repeated five-fold cross-validation and class-balanced batches to stabilize learning.
- The study compares a plain feedforward CNN with convolutional and pooling layers against a modern residual neural network.
- Experimental setup: Experiments tune layer count, filter counts, learning rate, pooling size, and dropout probability for each architecture.
- VoxCNN: The plain VoxCNN uses volumetric convolutional blocks, deconvolutional layers with batch normalization and dropout, and a softmax binary-classification output.
- ResNet: The ResNet architecture contains 21 layers with six VoxRes blocks, strided convolutions, pooling, a fully connected layer, and a softmax binary-classification output.
- Experimental setup: Performance is estimated with five-fold cross-validation using five different fold splits.
- Experimental setup: Each batch is constructed to include samples from every class because class-imbalanced batches can destabilize learning on the small dataset.
3. DATA
The experiments use preprocessed, spatially normalized ADNI structural MRI scans, selecting one image per subject to avoid information leakage. The resulting four-class dataset contains 231 3D images represented as 110 × 110 × 110 voxel tensors.
- The MRI subset was aligned, skull-stripped, spatially normalized, masked, and N3 corrected before use.
- Only the first image from each subject was selected to prevent possible information leaks from repeated scans.
- The dataset contains 231 images across AD, LMCI, EMCI, and NC classes.The class counts are 50 AD, 43 LMCI, 77 EMCI, and 61 NC images.
- Each MRI is stored as a 3D voxel-intensity tensor with shape 110 × 110 × 110.
4. RESULTS
The proposed networks classify Alzheimer’s disease versus normal controls accurately but struggle with mild cognitive impairment, while validation ROC AUC improves before plateauing after epoch 50. Attention maps highlight regions associated with Alzheimer’s disease.
- The networks accurately distinguish Alzheimer’s disease from normal controls but struggle to separate Alzheimer’s disease from early and late mild cognitive impairment.Both architectures show similar results within a standard deviation.
- Table 1 reports classification ROC AUC and accuracy as mean ± standard deviation across the binary tasks.
- Validation ROC AUC steadily improves during training and plateaus after the 50th epoch.
- The attention heatmap for a normal-cohort MRI identifies visually blurry regions whose highest-attention spots appear to overlap the hippocampus and ventricles.Attention is estimated by obstructing image regions with a 7×7×7 box and measuring the resulting drop in output probability.
5. CONCLUSION
The paper proposes deep residual and plain 3D convolutional networks for brain MRI classification, achieving performance comparable to previous approaches without handcrafted feature generation. Future work targets unaligned and unskullstripped images to enable a more streamlined analysis pipeline.
- The proposed deep 3D convolutional networks achieve performance comparable to previously used approaches on ADNI MRI classification.
- Their stated advantages are ease of use and no need for handcrafted feature generation.
- Future work aims to obtain similar or better results without alignment and skull-stripping preprocessing.
- Such preprocessing robustness could support one-step analysis instead of the multistep MRI pipelines currently described.