Computer-aided diagnosis of pulmonary nodules on CT scans: segmentation and classification using 3D active contours

Abstract

We are developing a computer-aided diagnosis (CAD) system to classify malignant and benign lung nodules found on CT scans. A fully automated system was designed to segment the nodule from its surrounding structured background in a local volume of interest (VOI) and to extract image features for classification. Image segmentation was performed with a three-dimensional (3D) active contour (AC) method. A data set of 96 lung nodules (44 malignant, 52 benign) from 58 patients was used in this study. The 3D AC model is based on two-dimensional AC with the addition of three new energy components to take advantage of 3D information: (1) 3D gradient, which guides the active contour to seek the object surface, (2) 3D curvature, which imposes a smoothness constraint in the z direction, and (3) mask energy, which penalizes contours that grow beyond the pleura or thoracic wall. The search for the best energy weights in the 3D AC model was guided by a simplex optimization method. Morphological and gray-level features were extracted from the segmented nodule. The rubber band straightening transform (RBST) was applied to the shell of voxels surrounding the nodule. Texture features based on run-length statistics were extracted from the RBST image. A linear discriminant analysis classifier with stepwise feature selection was designed using a second simplex optimization to select the most effective features. Leave-one-case-out resampling was used to train and test the CAD system. The system achieved a test area under the receiver operating characteristic curve (A(z)) of 0.83 +/- 0.04. Our preliminary results indicate that use of the 3D AC model and the 3D texture features surrounding the nodule is a promising approach to the segmentation and classification of lung nodules with CAD. The segmentation performance of the 3D AC model trained with our data set was evaluated with 23 nodules available in the Lung Image Database Consortium (LIDC). The lung nodule volumes segmented by the 3D AC model for best classification were generally larger than those outlined by the LIDC radiologists using visual judgment of nodule boundaries.

Fig. 1

Distribution of the longest diameters of the lung nodules in the data set, as measured by experienced thoracic radiologists on the axial slices of the CT examinations.

Fig. 2

Confidence ratings for the likelihood of a nodule being malignant (1 = most likely benign, 5 = most likely malignant) by experienced thoracic radiologists.

Fig. 3

The vertices of the polygon and positions used in the active contour model.

Fig. 4

An example demonstrating the correction of lung segmentation from the pleural surface. From left to right: the initial pleural boundary, indentation created along the lung boundary after local refinement, and corrected lung boundary after indentation is filled.

Fig. 5

The rubber band straightening transform (RBST). Top Left: A ROI containing a nodule. Top Right: The active contour boundary from which the RBST image is extracted. Bottom: The RBST image that will be Sobel-filtered, from which run-length statistics may be extracted. The black area of the RBST image corresponds to the pixels where the chest wall is masked out.

Fig. 6

Flow chart showing the simplex optimization process for selection of weights in the 3D AC model and classifier design.

Fig. 7

ROC curves comparing the different results for optimization using A_z as FOM. Computer A_z = 0.83, Rad features A_z=0.80, Rad likelihood A_z = 0.84.

Fig. 8

Test discriminant scores of lung nodules from the leave-one-case-out segmentation training and testing method.

Fig. 9

Overlap measures (a) and volume percentage errors (b) at different pmap thresholds for testing of the 3D AC segmentation using the 23 LIDC nodules. The error bars indicate 1 s.d. from the average (only one side shown for clarity). Two overlap measures are shown: Overlap₁(A,L) relative to the gold standard volume and Overlap₂(A,L) relative to the union of the segmented volume and the gold standard volume. Note the increasing volume error as the pmap threshold increases because the LIDC-defined nodule volume decreases with increasing pmap threshold values. A pmap value of 1000 means the intersection of all 18 LIDC manually and semiautomatically drawn contours by radiologists. Eight of the nodules contain no voxels in the intersection at pmap of 1000 so that the average and the median were calculated from the remaining 15 nodules.

Fig. 10

Percentage of volume error relative to the volume (in log scale) enclosed within the contour defined by a pmap threshold of 500 for each of the 23 LIDC test nodules. One small juxta-vascular nodule had a volume error of 743% because the blood vessel was erraneously segmented.

Fig. 11

Representative slices (not to scale) from difficult-to-segment LIDC nodules: (a) small, faint juxtavascular nodule (longest diameter 4.32 mm), (b) small nodule (longest diameter 4.98 mm), and (c) juxtavascular (longest diameter 11.92 mm) low contrast nodule in noisy image.

Fig. 12

An example of a nodule which was difficult to segment because it was embedded in thick blood vessels, leading to inaccurate classification. (a) axial slice through the nodule, and (b) 3D volume containing the nodule.

Fig. 13

The performance metrics TPF, FPR, FNF, and NOVR derived from the average Overlap₁(A,L) and V_err measures shown in Fig. 9.