04 - Basic Segmentation

Question 1

Segmentation approaches

Answer 1

• experimentalist = pb-driven

- computer vision = result-driven

Question 2

Thresholding

Answer 2

• use histogram to guess where the delimiting value is
• quantify against ground truth
• – sensitivity = TP/TP+FN = recall = TP rate
• – FR rate = FP/FP+TN
• – specificity = TN/TN+FP
• – precision = TP/TP+FP

Question 3

ROC

Answer 3

Receiver Operating Characteristic

• compute TPR and FPR at different threshold values
• use area under curve (AUC) to compare, the highest closest to 1 is the best

Question 4

Morphology

Answer 4

• morphological operation = usage of neighborhood voxels’ info to improve the result of thresholding
• assume noise and artefacts are less spatially correlated than the real values of nearby voxels
• erosion: 1 neighbor is 0 => becomes 0
• dilation: 1 neighbor is 1 => becomes 1
• opening: erosion then dilation (remove small obj/connections)
• closing: dilation then erosion (connect close objects)

Question 5

Segmentation pitfalls

Answer 5

• partial volume effect => discretization of the volume decrease representativity
• rescaling => apparent volume fraction chg when changing resolution