Source-linked AI summary
Resting state fMRI functional connectivity-based classification using a convolutional neural network architecture
Regina Meszlényi, Krisztian Buza, Zoltán Vidnyánszky
TL;DR
Small MRI samples pose an overfitting challenge for deep connectome classifiers. The paper proposes a connectivity-fingerprint CCNN and tests it on simulated data and aMCI classification. CCNN outperformed simple and deep neural networks, and combining connectivity metrics improved its classification performance.
Problem
Small MRI datasets create a challenge for deep connectome classifiers because large networks can overfit when training examples are limited.
Source: §1 · p. 2Method
The CCNN applies shared convolutional filters to ROI connectivity fingerprints and can combine different connectivity metrics as input channels.
Source: §1 · p. 3Results
CCNN systematically outperformed simple and deep neural networks on the evaluated tasks, and its combined-metric models performed best.
Source: §4 · pp. 12–13Takeaways & Limitations
The CCNN can use multiple functional-connectivity descriptors, a flexibility the authors identify as promising for future connectome research.
Source: §9.4 · p. 23Hyperparameter selection is difficult because deep models take time to train, limiting how thoroughly parameters can be explored.
Source: §4 · p. 14
Abstract
from arXiv · showhide
Machine learning techniques have become increasingly popular in the field of resting state fMRI (functional magnetic resonance imaging) network based classification. However, the application of convolutional networks has been proposed only very recently and has remained largely unexplored. In this paper we describe a convolutional neural network architecture for functional connectome classification called connectome-convolutional neural network (CCNN). Our results on simulated datasets and a publicly available dataset for amnestic mild cognitive impairment classification demonstrate that our CCNN model can efficiently distinguish between subject groups. We also show that the connectome-convolutional network is capable to combine information from diverse functional connectivity metrics and that models using a combination of different connectivity descriptors are able to outperform classifiers using only one metric. From this flexibility follows that our proposed CCNN model can be easily adapted to a wide range of connectome based classification or regression tasks, by varying which connectivity descriptor combinations are used to train the network.
1 Introduction
The paper introduces CCNN, a convolutional architecture for functional-connectome classification, and evaluates it on simulated data and a public aMCI dataset. Its design addresses connectome-specific structure and combines complementary connectivity measures.
1 Introduction
- Small MRI cohorts raise overfitting concerns for large deep networks, motivating relatively simple convolutional architectures for connectome classification.The paper notes MRI datasets typically contain dozens to a few thousand measurements, versus 1.2 million ImageNet instances.Source: §1 · p. 2
- Connectome matrices resemble images, but ROI ordering lacks the meaningful local neighborhoods and shift invariance that standard image convolutions assume.Because brain graphs are typically fully connected, the authors argue filters should account for the graph structure underlying the matrix.Source: §1 · pp. 2–3
- Although deep learning has been applied to fMRI and connectomes, convolutional connectome work remained early, and prior structural-connectome filters did not combine metric matrices as channels.The paper proposes treating matrices for different connectivity metrics like color channels, preserving correspondence between ROI pairs.Source: §1 · pp. 2–3
- CCNN shares filter weights across whole ROI connectivity fingerprints, producing one output per ROI with as many trainable weights as ROIs.The authors contrast this with an edge-to-node plus node-to-graph filter combination that would require three times as many weights.Source: §1 · p. 3
- Correlation-based connectivity can vary with brain state, while DTW distance and warping path length capture complementary connectivity properties for classification.The study examines combining these measures and compares them with correlation and the two DTW-derived measures individually.Source: §1 · pp. 3–4
- CCNN outperformed simple and deep neural networks on the evaluated tasks, with its best-performing models combining connectivity metrics.The study tests feasibility on simulated data and evaluates aMCI classification on a public dataset.Source: §1 · p. 4
2 Materials and methods
The methods combine an aMCI dataset and simulated connectomes with ROI-based connectivity measures and cross-validated classification. The proposed CCNN is evaluated alongside random, simple neural-network, and deep neural-network baselines.
2.1 Amnestic mild cognitive impairment dataset
- The aMCI data comprise 49 older adults, including 25 diagnosed with aMCI, with 146 resting-state measurements; preprocessing included motion correction, normalization, nuisance regression, and 0.009–0.08 Hz filtering.Measurements were collected at least twice per subject; preprocessing used SPM12 and custom MATLAB scripts.Source: §2.1 · p. 4, §2.2 · pp. 4–5
2.3 Functional connectivity calculation
- The Willard atlas supplied 499 ROI time series, yielding 124251 independent pairwise features characterized by correlation, DTW distance, and warping path length.The DTW measures are described as functional-connectivity metrics.Source: §2.3 · p. 5
2.4 Simulated dataset
- The simulated datasets replaced one, five, or ten ROI rows and columns in a healthy connectome with aMCI values, then varied added noise across ten levels.Each dataset contained 150 instances, and the three modification levels crossed with ten noise levels produced 30 datasets.Source: §2.4 · p. 5
2.5 Classification
- Evaluation used subject-level seven-fold cross-validation for aMCI and ten-fold cross-validation for simulation, with accuracy and AUC as performance measures.The random-classifier significance baselines were set using the 95th percentile of a cumulative binomial distribution.Source: §2.5 · pp. 5–6, §2.5.2 · p. 7
2.5.1 Connectome-convolutional neural network
- CCNN applies line-wise 1×499 and column-wise 499×1 convolutions, followed by a 96-neuron hidden layer and two class outputs.Its 64 first-layer filters extract features per ROI; 128 second-layer filters produce the feature vector, and two metric matrices can be input as channels with less than a one-percent increase in model size.Source: §2.5.1 · p. 6
2.5.3 Simple neural network classifier
- The simple neural-network baseline used 124251 inputs, one 128-neuron hidden layer, and two output neurons; its single-metric model had 15,904,384 trainable weights plus 130 biases.Combined inputs nearly doubled the weight count.Source: §2.5.3 · p. 7
2.5.4 Deep neural network classifier
- The deep neural-network baseline used hidden layers of 128 and 96 neurons, with ReLU, dropout, and Adam; its single-metric model had 15,916,608 trainable weights plus 226 biases.Figure 1 contrasts the simple, deep, and proposed CCNN architectures.Source: §2.5.4 · p. 7, Figure 1 · p. 8
3 Results
Across simulated data and aMCI classification, CCNN performed strongly, with its best aMCI result coming from combined DTW distance and warping path length features. Its learned weights also highlighted classification-relevant ROIs, with different ROIs emphasized by the two connectivity measures.
3 Results
- The results section reports accuracy and AUC and uses a binomial test, treating p-values below 0.05 as significant.The learned-weight analyses visualize features and ROIs influential to classification.Source: §3 · p. 8, §3.2 · p. 12
3.1 Classification of simulated data
- On simulated data, CCNN had the best overall performance and reached 100% accuracy at low noise levels, where the deep network could perform at or near chance.CCNN significantly outperformed the deep network in several cases.Source: §3.1 · pp. 8–9, Figure 2 · p. 9
- At noise level five, first-layer weights identified the altered ROI in the one-ROI dataset and recovered four of five altered ROIs in each of the five- and ten-ROI datasets.The authors caution that low first-layer weights do not establish that an ROI is unimportant to the full network.Source: §3.1 · pp. 9–10
3.2 Amnestic mild cognitive impairment classification
- For aMCI, only DTW-based measures exceeded the 56.85% random-classification threshold across the examined feature sets and networks.Among single-feature models, the DTW-distance CCNN performed best; CCNN’s 7% accuracy advantage over the simple network on path-length features was not significant (p=0.15).Source: §3.2 · p. 10, Table 1 · p. 10
- The combined-feature CCNN achieved the overall best aMCI classification performance, significantly outperforming the deep network (p=7.7e-4) and the path-length-only CCNN (p=0.031).Its performance did not differ significantly from the DTW-distance-only CCNN (p=0.092).Source: Table 1 · p. 10, §3.2 · p. 11
- In the combined aMCI model, the most influential ROIs for DTW distance and warping path length showed no significant overlap.The paper describes DTW distance as measuring connectivity strength and path length as connection stability; the two feature types highlighted different ROI sets.Source: §3.2 · pp. 11–12, Figure 3 · p. 11, Figure 4 · p. 12
4 Discussion
The CCNN outperformed simple and deep neural networks in aMCI classification, with its best accuracy from combining DTW distance and warping path length. Simulations also showed reduced overfitting and recovery of altered ROIs, while extremely small training sets remain a concern.
4 Discussion
- In simulations, CCNN systematically outperformed deep and simple neural architectures across noise and modification levels, was less prone to overfitting, and recovered altered ROIs.The recovered regions had the largest absolute weights in first-layer convolutional filters.Source: §4 · pp. 12–13
- SVM and LASSO did not outperform CCNN; the best LASSO classification was 2% less accurate than CCNN.Both traditional methods performed better than the deep neural network in the relatively small-sample setting.Source: §4 · p. 13
- In aMCI classification, influential regions included default-mode and executive-control network areas, as well as the hippocampus, basal ganglia, and amygdala.These regions aligned with prior research on resting-state connectivity abnormalities in aMCI.Source: §4 · p. 13
- DTW distance and warping path length capture complementary connectivity properties, helping explain why their combination performed best.Long warping paths can indicate unstable phase relationships even when DTW distance is relatively low.Source: §4 · pp. 13–14
- CCNN still requires careful regularization, and its effectiveness with fewer than a hundred measurements is debatable.Long training times also limit thorough hyperparameter tuning.Source: §4 · p. 14
- The best aMCI classification accuracy came from a CCNN combining DTW distance and warping path length, which outperformed deep and simple neural networks and correlation-based models.The combined DTW and path-length metrics performed significantly better than correlation coefficients.Source: §4 · pp. 13–14
6 Author Contributions
All coauthors contributed to the study.
6 Author Contributions
- All coauthors contributed to the study: RM, KB, and ZV.Source: §6 · p. 15
7 Funding
The paper reports support from the Hungarian Brain Research Program and additional support for Krisztian Buza.
7 Funding
- The study received a Hungarian Brain Research Program grant to Zoltán Vidnyánszky, and Krisztian Buza received NKFIH PD 111710 and János Bolyai Research Scholarship support.Source: §7 · p. 15
9 Supplementary Material - Dynamic Time Warping distance and warping path length
Dynamic Time Warping (DTW) measures the minimum cost of matching two time series while allowing elements to align flexibly. The warping path records that optimal matching, and its length relative to the diagonal reflects phase differences and their stability.
9 Supplementary Material - Dynamic Time Warping distance and warping path length
- DTW distance is the minimum cost of transforming one time series into another through element replacement and elongation steps.The method accumulates matching costs in a matrix whose dimensions correspond to the lengths of the two series.Source: §9 · p. 20
- Each DTW matrix entry adds the paired-value distance to the minimum cost among its permitted neighboring entries.The boundary entries use the available preceding cell; interior entries use the left, upper, or diagonal predecessor.Source: §9 · p. 20
- A warping window limits the maximum allowed time shift and restricts calculations to matrix entries near the main diagonal.The optimal warping path gives the element-by-element matching that yields the minimal DTW distance.Source: §9 · p. 20, Figure 1 · p. 21
- Warping path length is the path's length relative to the main diagonal, characterizing overall phase difference and the stability of the time-delay structure.Identical time series follow the diagonal, while phase differences add elongation steps.Source: §9 · p. 20
9.1 Results of the traditional machine learning methods
On the combined DTW distance and warping path length dataset, neither traditional classifier outperformed CCNN. The section reports linear SVM and LASSO comparisons using feature-selection settings intended to match the neural network’s 128 features.
9.1 Results of the traditional machine learning methods
- The traditional-method comparison used linear SVM with ANOVA F-test feature selection and LASSO.The paper describes both as methods able to handle dimensionality through feature selection.Source: §9.1 · p. 21
9.2 Results on the combined DTW distance and warping path length dataset
- Neither SVM nor LASSO outperformed CCNN on the combined DTW distance and warping path length dataset.Predictions from the tested hyper-parameter setups did not differ significantly (p>0.15 in all cases).Source: §9.2 · p. 22, Table 1 · p. 22, Table 2 · p. 22
9.2.1 SVM results
- The linear SVM experiments selected the best 128 connectivity features by ANOVA F-test and tested five values of complexity parameter C.Results are summarized in Supplementary Table 1.Source: §9.2.1 · p. 22, Table 1 · p. 22
9.2.2 LASSO results
- LASSO experiments tested six regularization values to select approximately 128 connectivity features.Supplementary Table 2 reports accuracy and the mean and standard deviation of selected feature counts across seven cross-validation folds.Source: §9.2.2 · p. 22, Table 2 · p. 22
9.3 Results on the single datasets
On the single-feature datasets, the traditional classifiers did not benefit from combining connectivity metrics as CCNN did. On combined DTW distance and path length, their best accuracies were lower than CCNN’s, but the differences were not significant.
9.3 Results on the single datasets
- SVM and LASSO single-dataset experiments tested whether traditional classifiers could use correlation, DTW distance, and warping path length separately.The experiments followed comparisons on the combined dataset, where CCNN performed considerably better than on any single-feature dataset.Source: §9.3 · p. 22
9.3.1 SVM Results
- Single-dataset SVM classification used the best 128 features by ANOVA F-test and C = 0.05.Results are summarized in Supplementary Table 3.Source: §9.3.1 · p. 22, Table 3 · p. 22
9.3.2 LASSO results
- Single-dataset LASSO classification used λ = 0.0008, with results summarized in Supplementary Table 4.Source: §9.3.2 · p. 23, Table 4 · p. 23
9.4 Discussion of the results of the traditional algorithms
- On combined DTW distance and path length, the best SVM accuracy was 5.5% lower and the best LASSO accuracy was 2% lower than CCNN, with neither difference significant (p>0.1).The authors caution that testing different hyper-parameters may make these test-performance estimates over-optimistic.Source: §9.4 · p. 23
- SVM accuracy on single DTW was 0.6% higher than on combined data, while LASSO tied between single path length and combined data.The authors conclude traditional algorithms did not gain performance from combining metrics, unlike CCNN.Source: §9.4 · p. 23
9.5 Supplementary Figure
Supplementary Figure 2 summarizes first-layer CCNN weights for simulated datasets with one, five, or ten modified ROIs. Peaks in the ROI summaries overlap the modified regions to varying degrees.
9.5 Supplementary Figure
- For one modified ROI, the sole highest ROI-weight peak matched the altered ROI; with five and ten modified ROIs, four peaks overlapped the altered regions in each case.The figure also shows 64×499 first-layer weight maps and summaries across filters and ROIs.Source: Figure 2 · p. 24
9.6 Supplementary Material - References
This supplementary section lists cited works spanning machine learning for neuroimaging, sparse classification, and functional-connectivity analysis.
9.6 Supplementary Material - References
- The references include work on sparse logistic regression and sparse network-based models for fMRI classification.Source: §9.6 · p. 25
- The bibliography includes references on dynamic time warping and functional-connectivity analysis.Source: §9.6 · p. 25
- Other listed works address time-series representations and deep learning.Source: §9.6 · p. 25