# False-Positive Region Reduction

Lymph node candidate region images contain many false-positive regions. A typical method of false-positive reduction is based on feature analysis. An example of a feature value set for false-positive region reduction is summarized in Table 3.2. A machine learning approach, such as a support vector machine, AdaBoost, or an artificial neural network can be used for this purpose. The typical procedure of falsepositive reduction based on a machine learning approach consists of two steps: (a) the training step and (b) the learning step. In the training step, features of trueTable 3.2 Feature value list for false-positive reduction

Shape features (16 features) |
Ratio of Vc and SAc. Ratio of the surface area of inscribed sphere and SAc- Sphericity. Inscribed sphericity. Circumscribed sphericity. Ratio of major axis and minor axis. Length of major and miner axis. Cross-section area at the midpoint of the each axis of BB. Distance between the comer point of BB and the centroid of BB. Ratio of SAc and the stuface area of Sp . Overlapping volume between Rc and Sp and the ratio of that volume and the volume of Ratio of г |

Intensity feanues (9 features) |
Average value. Variance value. Maximum value. Mininnun value. Median value. Fust quartile. Third quartile. Kurtosis. Skewness |

Feature values of intensity on the major and minor axis (10 features) |
Average value. Variance value. Maximum value. Minimum value. Median value |

Feanire value using around region of Rc (52 feanues) |
Overlapping volume between R$, and the air region. Average value, variance value, maximum value, minimum value and median value of Rd, and R |

Rc : Candidate region.

*’c* : Volume of Rc.

SAc : Surface area of Rc.

*Sy :* Sphere that has same volume and same gravity point of Rc- l *у :* Radius of Sj/.

Rd; : Dilated region of Rc by using spherical structure element of / mm in radius. (/'= 2.3.4.5) Rs, : Region that subtracts Rc from Rd, (t = 2.3.4.5)

BB : Bounding box of Rc- positive and false-positive regions are computed using training datasets. A set of true-positive and false-positive regions is obtained by performing the lymph node candidate region detection process described in Sects. 3.4.2 and 3.4.3. A classifier is trained by features generated from training datasets. Figure 3.13 shows a flowchart of the false-positive reduction process using a support vector machine.

After the training process of the classifier, feature values of lymph node candidate regions are computed. These features are entered into the trained classifier to decide whether a candidate region is or is not a lymph node.