The Segmentation Model

The segmentation model uses a U-Net architecture.
The data was augmented by rotating every image by an angle of

Comparison of Categorical Accuracy between Model 1 and Model 2.

Comparison of Loss between Model 1 and Model 2.

Comparison of Model IOU between Model 1 and Model 2.

T1 image	Model V1 predictions	Model V2 predictions

*The area enclosed by the red line indicates the resulting masks

Blind Performance Metrics

Faculty

	Model V1(5718 Images)		Model V2(10953 Images)
Class	Background	CC	Background	CC
Accuracy	0.995280886	0.995280886	0.998905092	0.998905092
Sensitivity	0.997443187	0.850984659	0.999347435	0.964329514
Specificity	0.850984659	0.997443187	0.964329514	0.999347435
PPV	0.99776626	0.832985664	0.99954356	0.949762916
NPV	0.832985664	0.99776626	0.949762916	0.99954356
IOU	0.995220842	0.726950136	0.99889159	0.917528605

Quality Control

The output masks from the segmentation are used to perform quality control.
This improves the masks generated.

Model Performance

Models	Ensemble	Neural Network	Wide & Deep	XGBOOST
Accuracy	96	95.4	96	97
Recall	78	77	77	80

Biological Associations

The relationships between subjects’ biological demographic information, such as age, sex, and health outcome, and extracted CC shape metrics from clean segmentations were analyzed to evaluate which features are most biologically meaningful. Regression and classification models were trained to predict age and neuropsychological health status, respectively. Different feature selection techniques were performed for each association and common relevant features selected in all generally tended to be thickness, curve, and Witelson5 sectioning metrics.