ITCFN: Incomplete Triple-Modal Co-Attention Fusion Network for Mild Cognitive Impairment Conersion Prediction

This repository contains the implementation of the algorithm presented in the paper titled "Title of the Paper" for MCI conversion prediction using MRI, PET, and clinical data.

1. Paper Algorithm Flowchart

2. Environment Setup

We conducted our experiments using the PyTorch 2.0 framework, utilizing a single NVIDIA A100 80 GB GPU for computational efficiency. The model was trained from scratch over two distinct stages, each consisting of 200 epochs, with a batch size of 8 to effectively manage the data. We optimized the model parameters using the Adam algorithm, setting the learning rate to 0.0001 to ensure precise adjustments during training.

Installation

pip install -r requirements.txt

3. Training and Inference Code

4. Dataset Folder and Split Ratios

To ensure reproducible and comparable results, we employed 5-fold cross-validation in all experiments, validating the model's stability and generalization while maintaining a consistent random seed for data splitting.

The dataset for this study is obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI), specifically the ADNI-1 and ADNI-2 cohorts. To prevent duplication, subjects present in both datasets were removed from ADNI-2. We selected T1-weighted sMRI, FDG-PET, and clinical data, categorized into four groups: normal controls (NC), sMCI, pMCI, and AD. Demographic information of the dataset is shown in Table below. Additionally, PET data is missing for 82 pMCI and 95 sMCI cases in ADNI-1, and for 1 pMCI and 30 sMCI cases in ADNI-2.

Variable	ADNI1 - AD	ADNI1 - pMCI	ADNI1 - sMCI	ADNI1 - NC	ADNI2 - AD	ADNI2 - pMCI	ADNI2 - sMCI	ADNI2 - NC
Number (M/F)	88/83	90/61	136/72	103/104	89/67	43/38	156/125	132/165
Age	75.35±7.47	74.63±7.18	74.75±7.63	75.92±5.12	74.75±8.09	72.60±7.27	71.29±7.43	72.80±6.01
Education	14.64±3.19	15.66±2.92	15.61±3.11	15.91±2.87	15.72±2.75	16.29±2.55	16.31±2.61	16.61±2.5
CDR-SB	4.32±1.58	1.85±0.98	1.38±0.75	0.03±0.12	4.51±1.67	2.18±0.95	1.33±0.82	0.04±0.15
MMSE	23.23±2.03	26.59±1.7	27.33±1.77	29.14±0.98	23.12±2.07	27.1±1.82	28.21±1.63	28.99±1.26

5. Code Execution Example

Here is an example of how to run the training and inference code:

Training: bash

python train.py --data_path <path_to_data> --epochs 50 --batch_size 32 Inference: bash python inference.py --model_path <path_to_trained_model> --data_path <path_to_data>

6. Experimental Results and Visualizations

The results of our experiments are summarized in the table below:

Methods	ADNI1	Modality	ACC	SPE	SEN	AUC	F1	ADNI2	ACC	SPE	SEN	AUC	F1
ResNet		M+P	0.725	0.823	0.564	0.653	0.606		0.809	0.928	0.437	0.709	0.510
JSRL		M+P	0.582	0.779	0.354	0.571	0.580		0.650	0.590	0.720	0.694	0.602
HOPE		M	0.611	0.786	0.611	0.648	0.593		0.712	0.860	0.712	0.616	0.692
VAPL		M+C	0.630	0.564	0.693	0.635	0.651		0.835	0.843	0.745	0.865	0.671
HFBSurv		M+P+C	0.921	0.904	0.937	0.920	0.916		0.954	0.977	0.909	0.943	0.932
ITCFN (w/o MMG)		M+P+C	0.932	0.925	0.937	0.931	0.927		0.960	0.992	0.937	0.965	0.960
ITCFN (Ours)		M+P+C	0.947	0.949	0.944	0.946	0.944		0.954	0.980	0.906	0.943	0.931

Methods	ADNI1	ACC	SPE	SEN	AUC	F1	ADNI2	ACC	SPE	SEN	AUC	F1
None		0.889	0.904	0.877	0.904	0.895		0.941	0.963	0.918	0.932	0.934
MMG only		0.889	0.890	0.891	0.890	0.879		0.948	0.976	0.914	0.935	0.932
TCAF only		0.932	0.925	0.937	0.931	0.927		0.960	0.992	0.937	0.965	0.960
MMG+TCAF		0.947	0.949	0.944	0.946	0.944		0.954	0.980	0.906	0.943	0.931