Fig 1.
The flowchart of the proposed method.
Table 1.
The specification of raw MAT images.
Table 2.
The specification of scale normalized MAT images.
Fig 2.
An image and its image features.
(A) An image. (B)-(F) The 0-th to 4-th level wavelet coefficients of (A) (Wavelet-0, …, Wavelet-4). (G)-(K) The histogram of the 0-th to 4-th level wavelet coefficients of (A) (HoW-0, …, HoW-4).
Table 3.
The specification of scale normalized MAT images.
Table 4.
The dimensionality of the image features.
Fig 3.
The elapsed time of the image processing steps [msec.].
(A) The elapsed time of pre-processing. (B) The elapsed time of feature extraction.
Fig 4.
The t-SNE 2D embedding of the features at the resolution of 256 × 256.
(A) Image, (B) Wavelet-3, and (C) HoW-3. Red (x) and purple (+) symbols represent the features of negative and positive patches respectively.
Fig 5.
A comparison of HoW-0 and HoW-4 features of patch size 256 × 256.
(A)-(C) HoW-0 features of Negative data. (D)-(F) HoW-4 features of Negative data. (G)-(I) HoW-0 features of Positive data. (J)-(L) HoW-4 features of Positive data.
Fig 6.
Visualized confusion matrix and MCC for each feature of different patch sizes.
(A) Patch size 256 × 256. (B) Patch size 512 × 512.
Fig 7.
A comparison of HoW-0 and HoW-4 features from the test cases of patch size 256 × 256.
Representative (A) HoW-0 and (B) HoW-4 features of true negative (TN), false negative (FN), false positive (FP) and true positive (TP) cases.
Table 5.
The elapsed time for training per dataset [sec.].
Table 6.
The elapsed time for test per patch [msec.].