Table 3 Summary of the cross-site E4 and E5 experiments.

From: Discriminative Scale Learning (DiScrn): Applications to Prostate Cancer Detection from MRI and Needle Biopsies

Method

Feature scales

AUC

Feature extraction time per slice (seconds)

DiScrn

Gabor: 3 × 3, LBP: 5 × 5

E4: 0.656

16.60

Haralick: 5 × 5, PHOW: 7 × 7

Gabor: 3 × 3, LBP: 5 × 5

E5: 0.657

214.46

Haralick: 5 × 5, PHOW: 7 × 7

T-test

Gabor: 3 × 3, 5 × 5, 7 × 7

E4: 0.613

N/A

LBP: 3 × 3, 5 × 5, 7 × 7

Haralick: 3 × 3, 5 × 5, 7 × 7

E5: 0.618

N/A

PHOW: 3 × 3, 5 × 5, 7 × 7

AllScales

Gabor: 3 × 3, 5 × 5, 7 × 7

E4: 0.640

28.54

LBP: 3 × 3, 5 × 5, 7 × 7

Haralick: 3 × 3, 5 × 5, 7 × 7

E5: 0.603

631.11

PHOW: 3 × 3, 5 × 5, 7 × 7

Baseline

Intensity: 1 × 1

E4: 0.518

≈0.0

E5: 0.475

  1. DiScrn is compared with using all predefined feature scales, T-test features selected from all predefined feature scales, and a baseline scheme that uses raw pixel intensities. The evaluation metrics are the AUC and the feature extraction time per slice at the testing stage. The second column shows the feature scales used for feature extraction. Note that we did not report the run time of T-test for feature extraction at the testing stage. This is because we did the feature extractions in parallel across all scales. In practice, feature selection can save the extraction time.
Back to article page