sd3004 / spatio-temporal-fmri-analysis

Spatio-Temporal Deep Learning for fMRI analysis

Dataset

Tensorboard

To visualise experiments logs in tensorbard run the following line:

tensorboard --logdir='./logs/'

Docker

docker build ./config

docker run -it --rm --env CUDA_VISIBLE_DEVICE=0 --gpus all -v /home/sd20/workspace:/workspace -v /home/sd20/workspace/data:/data/ --workdir=/workspace stgcn

docker tag 267a0d195121 stgcn

Useful Links

ST-GCN - Gadgil et al 2020, Spatio-Temporal Graph Convolution for Functional MRI Analysis

MS-G3D - Liu et al 2020, Disentangling and Unifying Graph Convolutions for Skeleton-Based Action Recognition

Results

Sex classification

Multi-layer Perceptron (MLP) classifier

Model	Data	Input Features	Architecture	Train-accuracy	Validation-accuracy	Remarks
ST-GCN	cov matrix - 22 ROIs	253	(64,64,1)	0.828	0.752	5-folds average, SGD 1e-2
ours	ICA15	105	(64,64,1)	1.00	0.847	dropout 0.5, Adam 1e-4
ours	ICA25	300	(64,64,1)	1.00	0.835	same
ours	ICA50	1225	(64,64,1)	1.00	0.902	same
ours	ICA100	4950	(64,64,1)	1.00	0.961	same
ours	ICA200	19900	(64,64,1)	1.00	0.957	same
ours	ICA300	44850	(64,64,1)	1.00	0.968	same

---

Graph Convolution Networks (GCN) classifiers

All results are obtained following a 5-fold cross validation

Model	Data	Accuracy % (paper)	Batch size	Iterations	Time Window	Remarks
ST-GCN	22 ROIs	81.8 (83.7)	5-folds average, SGD 1e-2	xxx	128	xxx
MS-G3D	22 ROIs	84.7	dropout 0.5, Adam 1e-4	xxx	50	xxx
MS-G3D	22 ROIs	83.0	dropout 0.0, Adam 1e-3	128	5k	scale 2, scale 2

Model	Data	Accuracy %	Batch size	Iterations	Time Window	Remarks
ST-GCN	Nodes TS - 15	79.8	512	10k	50	Adam, 1e-3
ST-GCN	Nodes TS - 15	75.9	512	2k	50	Adam, 1e-3
ST-GCN	Nodes TS - 15	76.7	512	2k	75	Adam, 1e-3
ST-GCN	Nodes TS - 15	76.6	512	2k	100	Adam, 1e-3
ST-GCN	Nodes TS - 25	82.1	512	10k	50	Adam, 1e-3
ST-GCN	Nodes TS - 25	78.3	512	2k	50	Adam, 1e-3
ST-GCN	Nodes TS - 25	79.8	512	2k	75	Adam, 1e-3
ST-GCN	Nodes TS - 25	77.8	512	2k	100	Adam, 1e-3
ST-GCN	Nodes TS - 25	76.5	512	2k	128	Adam, 1e-3
ST-GCN	Nodes TS - 50	86.5	512	10k	50	Adam, 1e-3
ST-GCN	Nodes TS - 50	82.4	512	2k	50	Adam, 1e-3
ST-GCN	Nodes TS - 50	81.6	512	2k	75	Adam, 1e-3
ST-GCN	Nodes TS - 50	79.6	512	2k	100	Adam, 1e-3
ST-GCN	Nodes TS - 100	82.6	256	2k	50	Adam, 1e-3
ST-GCN	Nodes TS - 200	89.2	256	2k	50	Adam, 1e-3
ST-GCN	Nodes TS - 200	91.5	256	10k	50	Adam, 1e-3
ST-GCN	Nodes TS - 300	87.2	128	2k	50	Adam, 1e-3

Model	Data	Accuracy %	Batch size	Iterations	Time Window	Remarks
MS-G3D	Nodes TS - 15	59.1	256	2k	10	Adam, 1e-3
MS-G3D	Nodes TS - 15	80.0	256	2k	50	Adam, 1e-3
MS-G3D	Nodes TS - 15	80.1	256	2k	75	Adam, 1e-3
MS-G3D	Nodes TS - 15	81.5	128	2k	100	Adam, 1e-3
MS-G3D	Nodes TS - 25	84.7	256	2k	50	Adam, 1e-3
MS-G3D	Nodes TS - 25	85.2	128	2k	75	Adam, 1e-3
MS-G3D	Nodes TS - 25	85.1	64	2k	100	Adam, 1e-3
MS-G3D	Nodes TS - 25	86.1	64	5k	100	Adam, 1e-3
MS-G3D	Nodes TS - 25	85.9	64	2k	128	Adam, 1e-3
MS-G3D	Nodes TS - 25	84.3	256	2k	50	Adam, 1e-3, scale g3d 4, scale gcn 4
MS-G3D	Nodes TS - 25	84.6	256	2k	50	Adam, 1e-3, scale g3d 1, scale gcn 1
MS-G3D	Nodes TS - 50	89.5	64	2k	50	Adam, 1e-3, scale g3d 8, scale gcn 18
MS-G3D	Nodes TS - 50	89.3	64	2k	75	Adam, 1e-3
MS-G3D	Nodes TS - 50	89.7	32	2k	100	Adam, 1e-3
MS-G3D	Nodes TS - 50	90.9	32	10k	100	Adam, 1e-3
MS-G3D	Nodes TS - 50	87.9	64	2k	50	Adam, 1e-3, scale g3d 4, scale gcn 1
MS-G3D	Nodes TS - 50	89	64	2k	50	Adam, 1e-3, scale g3d 1, scale gcn 1
MS-G3D	Nodes TS - 50	87.5	64	2k	50	Adam, 1e-3, scale g3d 1, scale gcn 1 , ws=1
MS-G3D	Nodes TS - 100	92.2	256	2k	50	Adam, 1e-3
MS-G3D	Nodes TS - 100	93.9	32	10k	50	Adam, 1e-3
MS-G3D-light	Nodes TS - 200	94.4	128	2k	50	Adam, 1e-3, scale 1

Fluid intelligence prediction

	Train-accuracy	Validation-accuracy
ICA15	xxx	xxx
ICA25	xxx	xxx
ICA50	xxx	xxx
ICA100	xxx	xxx
ICA200	xxx	xxx
ICA300	xxx	xxx

Model	Data	Correlation %	Batch size	Iterations	Time Window	Remarks
MS-G3D	Nodes TS - 15	28.6	256	2k	100	Adam, 1e-3, dropout 0, scale 2
MS-G3D	Nodes TS - 15	24.1	256	2k	75	Adam, 1e-3, dropout 0, scale 2
MS-G3D	Nodes TS - 15	26.7	512	2k	50	Adam, 1e-3, dropout 0, scale 2
MS-G3D	Nodes TS - 25	30.6	64	2k	128	Adam, 1e-3, scale 8
MS-G3D	Nodes TS - 25	28.6	128	2k	100	Adam, 1e-3, dropout 0, scale 2
MS-G3D	Nodes TS - 25	28.0	256	2k	75	Adam, 1e-3, dropout 0, scale 2
MS-G3D	Nodes TS - 25	31.3	256	2k	50	Adam, 1e-3, dropout 0, scale 2
MS-G3D	Nodes TS - 25	0.0714	256	2k	50	new
MS-G3D	Nodes TS - 50	32.5	64	2k	50	Adam, 1e-3, dropout 0.5, scale 8
MS-G3D	Nodes TS - 50	30.7	128	2k	50	Adam, 1e-3, dropout 0, scale 2
MS-G3D	Nodes TS - 50	28.5	128	2k	75	Adam, 1e-3, dropout 0, scale 2
MS-G3D	Nodes TS - 50	30.2	64	2k	100	Adam, 1e-3, dropout 0, scale 2
MS-G3D	Nodes TS - 100	31.7	32	2k	50	Adam, 1e-3, scale 8
MS-G3D	Nodes TS - 100	31.3	64	2k	50	Adam, 1e-3, dropout 0, scale 2
MS-G3D-light	Nodes TS - 200	32.4	16	2k	50	Adam, 1e-3, scale 1