Computer-aided detection of pulmonary nodules in low-dose CT

A computer-aided detection (CAD) system for the identification of pulmonary nodules in lowdose multi-detector helical CT images with 1.25-mm slice thickness is being developed in the framework of the INFN-supported MAGIC-5 Italian project. The basic modules of our lung-CAD system, a dotenhancement filter for nodule candidate selection and a voxel-based neural classifier for false-positive finding reduction, are described. Preliminary results obtained on the so-far collected database of lung CT scans are discussed. a local maximum detector) is then applied to the filter output to detect the filtered-signal peaks. Figure 2. Some examples of false positive findings generated by the dot-enhancement filter. Since most FP findings are provided by crossings between blood vessels (see Fig. 2), we attempted to reduce the amount of FP/scan by developing a procedure which we called voxel-based neural approach (VBNA). According to that method, each voxel of a region of interest (ROI) is characterized by the grey level intensity values of its neighbors (see Fig. 3). We developed, implemented and compared two different VBNA procedures. In the first, the CT values of the voxels in a 3D neighborhood of each voxel of a ROI are rolled down into vectors of features (147 features) to be analyzed by a neural classifier. In the second procedure (Gori, I. & Mattiuzzi, M. 2005), 6 additional features constituted by the eigenvalues of the gradient and the Hessian matrices are computed for each voxel and encoded to the feature vectors (153 features). A feed-forward neural network is implemented at this stage to assign each voxel either to the nodule or normal tissue target class. A candidate nodule is then characterized as “CAD nodule” if the number of pixels within its ROI tagged as “nodule” by the neural classifier is above some relative threshold. A free response receiver operating characteristic (FROC) curve for our CAD system can therefore be evaluated at different threshold levels. Figure 3. Voxel-based neural approach to false-positive reduction. 3 DATA ANALYSIS AND RESULTS The CAD system was developed and tested on a dataset of low-dose (screening setting: 140 kV, 70÷80 mA) CT scans with reconstructed slice thickness of 1.25 mm. The scans were collected and annotated by experienced radiologists in the framework of the screening trial being conducted in Italy (Italung-CT). The database available for this study consists of 14 scans, containing 24 internal nodules. Each scan is a sequence of about 300 slices stored in the DICOM (Digital Imaging and COmmunications in Medicine) format. First of all, the lung volume is segmented out of the whole 3D data array by means of a purposely built segmentation algorithm that identifies the internal region of the lung (Antonelli et al. 2005). The 3D dot-enhancement filter applied to the selected lung regions shows a very high sensitivity. In particular, the lists generated by the peak-detector algorithm for all CT are empirically truncated so to include all annotated nodules. According to this procedure, a 100% sensitivity to internal nodules is obtained at a maximum (average) number of 54 (52.3) FP/scan. With respect to the VBNA procedure for FP reduction, the dataset was randomly partitioned into train and test sets; the performances of the trained neural networks were evaluated both on the test sets and on the whole dataset. In the first VBNA approach, 147 features, derived from a 2D region of 7x7 voxels for 3 consecutive slices with the voxel to be classified in the center, constitute each vector of the feature dataset. Two three-layer feed-forward neural networks with 147 input, were trained on two different random partitions of the dataset into train and test sets. The performances achieved in each trial for the correct classification of individual pixels are reported in Table 1, where the sensitivity and the specificity values obtained on the test sets, on the whole datasets and the average values on the two trials are shown. Table 1. VBNA with 147 features ______________________________________________ test train+test _______________ _______________ sens % spec % sens % spec % ______________________________________________