Ted W. Way

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

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.

Effect of CT scanning parameters on volumetric measurements of pulmonary nodules by 3D active contour segmentation: a phantom study

The purpose of this study is to investigate the effects of CT scanning and reconstruction parameters on automated segmentation and volumetric measurements of nodules in CT images. Phantom nodules of known sizes were used so that segmentation accuracy could be quantified in comparison to ground-truth volumes. Spherical nodules having 4.8, 9.5 and 16 mm diameters and 50 and 100 mg cc(-1) calcium contents were embedded in lung-tissue-simulating foam which was inserted in the thoracic cavity of a chest section phantom. CT scans of the phantom were acquired with a 16-slice scanner at various tube currents, pitches, fields-of-view and slice thicknesses. Scans were also taken using identical techniques either within the same day or five months apart for study of reproducibility. The phantom nodules were segmented with a three-dimensional active contour (3DAC) model that we previously developed for use on patient nodules. The percentage volume errors relative to the ground-truth volumes were estimated under the various imaging conditions. There was no statistically significant difference in volume error for repeated CT scans or scans taken with techniques where only pitch, field of view, or tube current (mA) were changed. However, the slice thickness significantly (p < 0.05) affected the volume error. Therefore, to evaluate nodule growth, consistent imaging conditions and high resolution should be used for acquisition of the serial CT scans, especially for smaller nodules. Understanding the effects of scanning and reconstruction parameters on volume measurements by 3DAC allows better interpretation of data and assessment of growth. Tracking nodule growth with computerized segmentation methods would reduce inter- and intraobserver variabilities.

Effect of finite sample size on feature selection and classification: A simulation study

PURPOSE The small number of samples available for training and testing is often the limiting factor in finding the most effective features and designing an optimal computer-aided diagnosis (CAD) system. Training on a limited set of samples introduces bias and variance in the performance of a CAD system relative to that trained with an infinite sample size. In this work, the authors conducted a simulation study to evaluate the performances of various combinations of classifiers and feature selection techniques and their dependence on the class distribution, dimensionality, and the training sample size. The understanding of these relationships will facilitate development of effective CAD systems under the constraint of limited available samples. METHODS Three feature selection techniques, the stepwise feature selection (SFS), sequential floating forward search (SFFS), and principal component analysis (PCA), and two commonly used classifiers, Fishers linear discriminant analysis (LDA) and support vector machine (SVM), were investigated. Samples were drawn from multidimensional feature spaces of multivariate Gaussian distributions with equal or unequal covariance matrices and unequal means, and with equal covariance matrices and unequal means estimated from a clinical data set. Classifier performance was quantified by the area under the receiver operating characteristic curve Az. The mean Az values obtained by resubstitution and hold-out methods were evaluated for training sample sizes ranging from 15 to 100 per class. The number of simulated features available for selection was chosen to be 50, 100, and 200. RESULTS It was found that the relative performance of the different combinations of classifier and feature selection method depends on the feature space distributions, the dimensionality, and the available training sample sizes. The LDA and SVM with radial kernel performed similarly for most of the conditions evaluated in this study, although the SVM classifier showed a slightly higher hold-out performance than LDA for some conditions and vice versa for other conditions. PCA was comparable to or better than SFS and SFFS for LDA at small samples sizes, but inferior for SVM with polynomial kernel. For the class distributions simulated from clinical data, PCA did not show advantages over the other two feature selection methods. Under this condition, the SVM with radial kernel performed better than the LDA when few training samples were available, while LDA performed better when a large number of training samples were available. CONCLUSIONS None of the investigated feature selection-classifier combinations provided consistently superior performance under the studied conditions for different sample sizes and feature space distributions. In general, the SFFS method was comparable to the SFS method while PCA may have an advantage for Gaussian feature spaces with unequal covariance matrices. The performance of the SVM with radial kernel was better than, or comparable to, that of the SVM with polynomial kernel under most conditions studied.

Characterization of masses in digital breast tomosynthesis: Comparison of machine learning in projection views and reconstructed slices

PURPOSE In digital breast tomosynthesis (DBT), quasi-three-dimensional (3D) structural information is reconstructed from a small number of 2D projection view (PV) mammograms acquired over a limited angular range. The authors developed preliminary computer-aided diagnosis (CADx) methods for classification of malignant and benign masses and compared the effectiveness of analyzing lesion characteristics in the reconstructed DBT slices and in the PVs. METHODS A data set of MLO view DBT of 99 patients containing 107 masses (56 malignant and 51 benign) was collected at the Massachusetts General Hospital with IRB approval. The DBTs were obtained with a GE prototype system which acquired 11 PVs over a 50 degree arc. The authors reconstructed the DBTs at 1 mm slice interval using a simultaneous algebraic reconstruction technique. The region of interest (ROI) containing the mass was marked by a radiologist in the DBT volume and the corresponding ROIs on the PVs were derived based on the imaging geometry. The subsequent processes were fully automated. For classification of masses using the DBT-slice approach, the mass on each slice was segmented by an active contour model initialized with adaptive k-means clustering. A spiculation likelihood map was generated by analysis of the gradient directions around the mass margin and spiculation features were extracted from the map. The rubber band straightening transform (RBST) was applied to a band of pixels around the segmented mass boundary. The RBST image was enhanced by Sobel filtering in the horizontal and vertical directions, from which run-length statistics texture features were extracted. Morphological features including those from the normalized radial length were designed to describe the mass shape. A feature space composed of the spiculation features, texture features, and morphological features extracted from the central slice alone and seven feature spaces obtained by averaging the corresponding features from three to 19 slices centered at the central slice were compared. For classification of masses using the PV approach, a feature extraction process similar to that described above for the DBT approach was performed on the ROIs from the individual PVs. Six feature spaces obtained from the central PV alone and by averaging the corresponding features from three to 11 PVs were formed. In each feature space for either the DBT-slice or the PV approach, a linear discriminant analysis classifier with stepwise feature selection was trained and tested using a two-loop leave-one-case-out resampling procedure. Simplex optimization was used to guide feature selection automatically within the training set in each leave-one-case-out cycle. The performance of the classifiers was evaluated by the area (Az) under the receiver operating characteristic curve. RESULTS The test Az values from the DBT-slice approach ranged from 0.87 +/- 0.03 to 0.93 +/- 0.02, while those from the PV approach ranged from 0.78 +/- 0.04 to 0.84 +/- 0.04. The highest test Az of 0.93 +/- 0.02 from the nine-DBT-slice feature space was significantly (p = 0.006) better than the highest test Az of 0.84 +/- 0.04 from the nine-PV feature space. CONCLUSION The features of breast lesions extracted from the DBT slices consistently provided higher classification accuracy than those extracted from the PV images.

The effect of nodule segmentation on the accuracy of computerized lung nodule detection on CT scans: comparison on a data set annotated by multiple radiologists

In computerized nodule detection systems on CT scans, many features that are useful for classifying whether a nodule candidate identified by prescreening is a true positive depend on the shape of the segmented object. We designed two segmentation algorithms for detailed delineation of the boundaries for nodule candidates. The first segmentation technique was a three-dimensional (3D) region-growing (RG) method which grew the object across multiple CT sections. The second technique was based on a 3D active contour (AC) model. A training set of 94 CT scans was used for algorithm design. An independent set of 62 scans, each read by multiple radiologists, was used for testing. Thirty-three scans were collected from patient files at the University of Michigan and 29 scans by the Lung Imaging Database Consortium (LIDC). In this study, we concentrated on the detection of internal lung nodules having a size ≥3 mm that were not pure ground-glass opacities. Of the lesions marked by one or multiple radiologists, 124 nodules satisfied these criteria and were considered true nodules. The performance of the detection system in the AC feature space, RG feature space, and the combined feature space were compared using free-response receiver operating curves (FROC). The FROC curve using the combined feature space was significantly higher than that using the RG feature space or the AC feature space alone (p=0.02 and 0.03, respectively). At a sensitivity of 70% for internal non-GGO nodules, the FP rates were 2.2, 2.2, and 1.5 per scan, respectively, for the RG, AC, and the combined methods. Our results indicate that the 3D AC algorithm can provide useful features to improve nodule detection on CT scans.

Quantitative CT of lung nodules: Dependence of calibration on patient body size, anatomic region, and calibration nodule size for single- and dual-energy techniques

Calcium concentration may be a useful feature for distinguishing benign from malignant lung nodules in computer-aided diagnosis. The calcium concentration can be estimated from the measured CT number of the nodule and a CT number vs calcium concentration calibration line that is derived from CT scans of two or more calcium reference standards. To account for CT number nonuniformity in the reconstruction field, such calibration lines may be obtained at multiple locations within lung regions in an anthropomorphic phantom. The authors performed a study to investigate the effects of patient body size, anatomic region, and calibration nodule size on the derived calibration lines at ten lung region positions using both single energy (SE) and dual energy (DE) CT techniques. Simulated spherical lung nodules of two concentrations (50 and 100 mg/cc CaCO3) were employed. Nodules of three different diameters (4.8, 9.5, and 16 mm) were scanned in a simulated thorax section representing the middle of the chest with large lung regions. The 4.8 and 9.5 mm nodules were also scanned in a section representing the upper chest with smaller lung regions. Fat rings were added to the peripheries of the phantoms to simulate larger patients. Scans were acquired on a GE-VCT scanner at 80, 120, and 140 kVp and were repeated three times for each condition. The average absolute CT number separations between the calibration lines were computed. In addition, under- or overestimates were determined when the calibration lines for one condition (e.g., small patient) were used to estimate the CaCO3 concentrations of nodules for a different condition (e.g., large patient). The authors demonstrated that, in general, DE is a more accurate method for estimating the calcium contents of lung nodules. The DE calibration lines within the lung field were less affected by patient body size, calibration nodule size, and nodule position than the SE calibration lines. Under- or overestimates in CaCO3 concentrations of nodules were also in general smaller in quantity with DE than with SE. However, because the slopes of the calibration lines for DE were about one-half the slopes for SE, the relative improvement in the concentration estimates for DE as compared to SE was about one-half the relative improvement in the separation between the calibration lines. Results in the middle of the chest thorax section with large lungs were nearly completely consistent with the above generalization. On the other hand, results in the upper-chest thorax section with smaller lungs and greater amounts of muscle and bone were mixed. A repeat of the entire study in the upper thorax section yielded similar mixed results. Most of the inconsistencies occurred for the 4.8 mm nodules and may be attributed to errors caused by beam hardening, volume averaging, and insufficient sampling. Targeted, higher resolution reconstructions of the smaller nodules, application of high atomic number filters to the high energy x-ray beam for improved spectral separation, and other future developments in DECT may alleviate these problems and further substantiate the superior accuracy of DECT in quantifying the calcium concentrations of lung nodules.

Computer-aided diagnosis for interval change analysis of lung nodule features in serial CT examinations

A CAD system was developed to extract and analyze features from corresponding malignant and benign lung nodules on temporal pairs of CT scans. The lung nodules on the current and prior CT scans were automatically segmented using a 3-dimensional (3D) active contour model. Three-dimensional run length statistics (RLS) texture features, 3D morphological and gray-level features were extracted from each nodule. In addition, 3D nodule profile features (PROF) that describe the gray level variation inside and outside the nodule surface were extracted by estimating the gradient magnitude values along the radial vectors from the nodule centroid to a band of voxels surrounding the nodule surface. Interval change features were calculated as the difference between the corresponding features extracted from the prior and the current scans of the same nodule. Stepwise feature selection with simplex optimization was used to select the best feature subset from the feature space that combined both the interval change features and features from the single current exam. A linear discriminant classifier was used to merge the selected features for classification of malignant and benign nodules. In this preliminary study, a data set of 103 nodule temporal pairs (39 malignant and 64 benign) was used. A leave-one-case-out resampling scheme was used for feature selection and classification. An average of 5 features was selected from the training subsets. The most frequently selected features included a difference PROF feature and 4 RLS features. The classifier achieved a test Az of 0.85±0.04. In comparison a classifier using features extracted from the current CT scans alone achieved a test Az of 0.78±0.05. This study indicates that our CAD system using interval change information is useful for classification of lung nodules on CT scans.

Digital tomosynthesis mammography : comparison of mass classification using 3D slices and 2D projection views

We are developing computer-aided diagnosis (CADx) methods for classification of masses on digital breast tomosynthesis mammograms (DBTs). A DBT data set containing 107 masses (56 malignant and 51 benign) collected at the Massachusetts General Hospital was used. The DBTs were obtained with a GE prototype system which acquired 11 projection views (PVs) over a 50-degree arc. We reconstructed the DBTs at 1-mm slice interval using a simultaneous algebraic reconstruction technique. The regions of interest (ROIs) containing the masses in the DBT volume and the corresponding ROIs on the PVs were identified. The mass on each slice or each PV was segmented by an active contour model. Spiculation measures, texture features, and morphological features were extracted from the segmented mass. Four feature spaces were formed: (1) features from the central DBT slice, (2) average features from 5 DBT slices centered at the central slice, (3) features from the central PV, and (4) average features from all 11 PVs. In each feature space, a linear discriminant analysis classifier with stepwise feature selection was trained and tested using a two loop leave-one-case-out procedure. The test Az of 0.91±0.03 from the 5-DBT-slice feature space was significantly (p=0.003) higher than that of 0.84±0.04 from the 1-DBT-slice feature space. The test Az of 0.83±0.04 from the 11-PV feature space was not significantly different (p=0.18) from that of 0.79±0.04 from the 1-PV feature space. The classification accuracy in the 5-DBT-slice feature space was significantly better (p=0.006) than that in the 11-PV feature space. The results demonstrate that the features of breast lesions extracted from the DBT slices may provide higher classification accuracy than those from the PV images.

SU‐FF‐I‐07: Single‐ and Dual‐Energy CT Calibration Lines for Assessing the Calcium Content of Lung Nodules: Effects of Patient Body and Lung Nodule Size

Purpose: To determine the concentration of Ca in lung nodules for a CAD technique, lung nodule calibration lines are being derived at locations throughout lung fields. A study was performed to investigate the effects of patient body and lung nodule size on derived calibration lines. Method and Materials: Simulated spherical lung nodules of two concentrations (50 and 100mg/cc CaCO 3 ) were employed. Three different diameter nodules (4.8mm, 9.5mm, 16mm) were scanned in a simulated thorax section “A” representing the middle of the chest with large lung regions. The 4.8mm and 9.5mm nodules were also scanned in section “B” representing the upper chest with smaller lung regions. Fat‐rings were added to the phantoms to simulate larger patients. Images were acquired on a GE‐VCT scanner at 80, 120 and 140kVp. The RMS CT♯ displacements between the calibration lines for phantoms with and without fat‐rings were compared. Results: Body‐size had a significant effect on the calibration lines for each single kVp technique. Mean RMS displacements for the 9.5mm nodules at 80, 120 and 140 kVp were 22+/−2, 19+/−3, and 18+/−2HU, respectively for phantom “A”, and 19+/−2, 14+/−1, and 14+/−1HU for “B”. Corresponding displacements for 80kVp–140kVp dual‐energy were much less: 5+/–1HU (“A”) and 6+/−2HU (“B”). Results similar to the 9.5mm were obtained for the 4.8 and 16mm nodules in phantom “A” However, in phantom “B”, the 4.8mm dual‐energy displacements (12+/−4 HU) were about as large as the single‐energy. The phantom “B” study was repeated, and the dual‐energy displacements for the 4.8mm nodules were slightly better (7+/− 4 HU) on one lung side but about as poor (10+/−7HU) on the other. Conclusion: Dual‐energy CTcalibration of the calcium concentration of lung nodules is less sensitive to patient body size than single‐energy calibration. However, the dual‐energy approach may not compensate for patient body size for smaller nodules.

SU‐EE‐A4‐01: CT Number Accuracy of Lung Nodules: Effect of Patient Body Size and Lung Size

Purpose: To investigate the effects of patient body size and lung size on the CT numbers of lung nodules measured with multi‐detector CTscanners and whether improved accuracy can be obtained with a dual‐energy technique. Method and Materials: Simulated lung nodules consisting of 9.5‐mm diameter spheres containing 50mg/cc and 100mg/cc CaCO3 in a water‐equivalent resin were scanned in two simulated thorax section phantoms with a GE VCT scanner. One phantom (A) represented the middle of the chest. It had large simulated lung regions and simulated ribs, heart and spine. The other (B) represented the upper chest. It had a much wider aspect ratio, smaller simulated lung regions, and simulated ribs, scapula, heart, and spine. Fat rings were added to the phantoms to simulate larger patients. Images were acquired on a GE VCT scanner with high‐resolution techniques (0.53:1 pitch, 0.625‐mm slice thickness and interval) at 80, 120 and 140kVp. Scans were repeated 3 times for reproducibility and analyzed using an automated technique. Results: Body size had a significant effect on the measured mean CT‐numbers of the nodules. For phantom‐A, adding fat rings decreased the overall average CT‐numbers of the 50mg/cc nodules at 120kVp by 15HU and those of the 100mg/cc nodules by 21HU. Corresponding reductions in phantom‐B were 9HU and 13HU. The dual‐energy approach (CT#80kVp‐CT#140kVp) reduces the variability, with a maximum difference of 4HU for all conditions. Lung size had a minimal effect with a maximum difference (nodule CT# phantom A ‐ nodule CT# phantom B) of 4.5 HU. Conclusion: Even with modern multi‐detector CTscanners, beam hardening and x‐ray scatter errors due to body size can result in underestimates of the true CT numbers of lung nodules. A dual‐energy approach compensates for these errors and should be considered especially if it can be implemented using a rapid kVp switching technique.