Table 5.

Comparison of classification performance over different experiments with ROI volume Feature.

	Accuracy	NC sen.	AD sen.	FTD sen.
PCA+SVM	82.28	85.94	85.44	67.14
ROI MLP	79.78	83.44	79.79	69.67
500 MLP	85.78	91.60	82.83	73.31
1000 MLP	85.41	90.03	84.91	73.07
2000 MLP	85.45	90.34	82.26	75.06
MDNN	85.97	91.05	83.88	74.20

The second column is the overall classification accuracy, while the third to fifth columns represent the sensitivity of NC, AD, and FTD, respectively. The second row represents the result using PCA+SVM with multi-scale volume size features, while the third row shows the classification performance of a single MLP with ROI-wise features. The fourth to sixth rows are the result of a single MLP with features extracted at different scales, i.e., 500, 1,000, and 2000 voxels per patch. The last row represents the MDNN result with multi-scale volume size features.