Wook-Jin Choi

Genetic programming-based feature transform and classification for the automatic detection of pulmonary nodules on computed tomography images

An effective automated pulmonary nodule detection system can assist radiologists in detecting lung abnormalities at an early stage. In this paper, we propose a novel pulmonary nodule detection system based on a genetic programming (GP)-based classifier. The proposed system consists of three steps. In the first step, the lung volume is segmented using thresholding and 3D-connected component labeling. In the second step, optimal multiple thresholding and rule-based pruning are applied to detect and segment nodule candidates. In this step, a set of features is extracted from the detected nodule candidates, and essential 3D and 2D features are subsequently selected. In the final step, a GP-based classifier (GPC) is trained and used to classify nodules and non-nodules. GP is suitable for detecting nodules because it is a flexible and powerful technique; as such, the GPC can optimally combine the selected features, mathematical functions, and random constants. Performance of the proposed system is then evaluated using the Lung Image Database Consortium (LIDC) database. As a result, it was found that the proposed method could significantly reduce the number of false positives in the nodule candidates, ultimately achieving a 94.1% sensitivity at 5.45 false positives per scan.

Automated pulmonary nodule detection based on three-dimensional shape-based feature descriptor

Computer-aided detection (CAD) can help radiologists to detect pulmonary nodules at an early stage. In pulmonary nodule CAD systems, feature extraction is very important for describing the characteristics of nodule candidates. In this paper, we propose a novel three-dimensional shape-based feature descriptor to detect pulmonary nodules in CT scans. After lung volume segmentation, nodule candidates are detected using multi-scale dot enhancement filtering in the segmented lung volume. Next, we extract feature descriptors from the detected nodule candidates, and these are refined using an iterative wall elimination method. Finally, a support vector machine-based classifier is trained to classify nodules and non-nodules. The performance of the proposed system is evaluated on Lung Image Database Consortium data. The proposed method significantly reduces the number of false positives in nodule candidates. This method achieves 97.5% sensitivity, with only 6.76 false positives per scan.

Automated Pulmonary Nodule Detection System in Computed Tomography Images: A Hierarchical Block Classification Approach

A computer-aided detection (CAD) system is helpful for radiologists to detect pulmonary nodules at an early stage. In this paper, we propose a novel pulmonary nodule detection method based on hierarchical block classification. The proposed CAD system consists of three steps. In the first step, input computed tomography images are split into three-dimensional block images, and we apply entropy analysis on the block images to select informative blocks. In the second step, the selected block images are segmented and adjusted for detecting nodule candidates. In the last step, we classify the nodule candidate images into nodules and non-nodules. We extract feature vectors of the objects in the selected blocks. Lastly, the support vector machine is applied to classify the extracted feature vectors. Performance of the proposed system is evaluated on the Lung Image Database Consortium database. The proposed method has reduced the false positives in the nodule candidates significantly. It achieved 95.28% sensitivity with only 2.27 false positives per scan.

PCA-based method for 3D shape recovery of microscopic objects from image focus using discrete cosine transform.

This article introduces a new algorithm for shape from focus (SFF) based on discrete cosine transform (DCT) and principal component analysis (PCA). DCT is applied on a small 3D neighborhood for each pixel in the image volume. Instead of summing all focus values in a window, AC parts of DCT are collected and then PCA is applied to transform this data into eigenspace. The first feature, containing maximum variation is employed to compute the depth. DCT and PCA are computationally intensive; however, the reduced data elements and algorithm iterations have made the new approach competitive and efficient. The performance of the proposed approach is compared with other methods by conducting experiments using image sequences of a synthetic and two microscopic objects. The evaluation is gauged on the basis of unimodality, monotonicity, and resolution of the focus curve. Two other global statistical metrics, root mean square error (RMSE) and correlation have also been applied for synthetic image sequence. Besides, noise sensitivity and computational complexity are also compared with other algorithms. Experimental results demonstrate the effectiveness and the robustness of the new method. Microsc. Res. Tech., 2008.

Adaptive region-growing with maximum curvature strategy for tumor segmentation in 18F-FDG PET

Accurate tumor segmentation in PET is crucial in many oncology applications. We developed an adaptive region-growing (ARG) algorithm with a maximum curvature strategy (ARG_MC) for tumor segmentation in PET. The ARG_MC repeatedly applied a confidence connected region-growing algorithm with increasing relaxing factor f. The optimal relaxing factor (ORF) was then determined at the transition point on the f-volume curve, where the volume just grew from the tumor into the surrounding normal tissues. The ARG_MC along with five widely used algorithms were tested on a phantom with 6 spheres at different signal to background ratios and on two clinic datasets including 20 patients with esophageal cancer and 11 patients with non-Hodgkin lymphoma (NHL). The ARG_MC did not require any phantom calibration or any a priori knowledge of the tumor or PET scanner. The identified ORF varied with tumor types (mean ORF  =  9.61, 3.78 and 2.55 respectively for the phantom, esophageal cancer, and NHL datasets), and varied from one tumor to another. For the phantom, the ARG_MC ranked the second in segmentation accuracy with an average Dice similarity index (DSI) of 0.86, only slightly worse than Daisnes adaptive thresholding method (DSI  =  0.87), which required phantom calibration. For both the esophageal cancer dataset and the NHL dataset, the ARG_MC had the highest accuracy with an average DSI of 0.87 and 0.84, respectively. The ARG_MC was robust to parameter settings and region of interest selection, and it did not depend on scanners, imaging protocols, or tumor types. Furthermore, the ARG_MC made no assumption about the tumor size or tumor uptake distribution, making it suitable for segmenting tumors with heterogeneous FDG uptake. In conclusion, the ARG_MC was accurate, robust and easy to use, it provides a highly potential tool for PET tumor segmentation in clinic.

Individually optimized contrast-enhanced 4D-CT for radiotherapy simulation in pancreatic ductal adenocarcinoma.

PURPOSE To develop an individually optimized contrast-enhanced (CE) 4D-computed tomography (CT) for radiotherapy simulation in pancreatic ductal adenocarcinomas (PDA). METHODS Ten PDA patients were enrolled. Each underwent three CT scans: a 4D-CT immediately following a CE 3D-CT and an individually optimized CE 4D-CT using test injection. Three physicians contoured the tumor and pancreatic tissues. Image quality scores, tumor volume, motion, tumor-to-pancreas contrast, and contrast-to-noise ratio (CNR) were compared in the three CTs. Interobserver variations were also evaluated in contouring the tumor using simultaneous truth and performance level estimation. RESULTS Average image quality scores for CE 3D-CT and CE 4D-CT were comparable (4.0 and 3.8, respectively; P = 0.082), and both were significantly better than that for 4D-CT (2.6, P < 0.001). Tumor-to-pancreas contrast results were comparable in CE 3D-CT and CE 4D-CT (15.5 and 16.7 Hounsfield units (HU), respectively; P = 0.21), and the latter was significantly higher than in 4D-CT (9.2 HU, P = 0.001). Image noise in CE 3D-CT (12.5 HU) was significantly lower than in CE 4D-CT (22.1 HU, P = 0.013) and 4D-CT (19.4 HU, P = 0.009). CNRs were comparable in CE 3D-CT and CE 4D-CT (1.4 and 0.8, respectively; P = 0.42), and both were significantly better in 4D-CT (0.6, P = 0.008 and 0.014). Mean tumor volumes were significantly smaller in CE 3D-CT (29.8 cm3, P = 0.03) and CE 4D-CT (22.8 cm3, P = 0.01) than in 4D-CT (42.0 cm3). Mean tumor motion was comparable in 4D-CT and CE 4D-CT (7.2 and 6.2 mm, P = 0.17). Interobserver variations were comparable in CE 3D-CT and CE 4D-CT (Jaccard index 66.0% and 61.9%, respectively) and were worse for 4D-CT (55.6%) than CE 3D-CT. CONCLUSIONS CE 4D-CT demonstrated characteristics comparable to CE 3D-CT, with high potential for simultaneously delineating the tumor and quantifying tumor motion with a single scan.

DCT and PCA Based Method for Shape from Focus

Discrete Cosine Transform (DCT) and Principal Component Analysis (PCA) are widely used in computer vision applications. In this paper, we introduce a new SFF method based on DCT and PCA. Contrary to computing focus quality locally by summing all values in a 2D or 3D window obtained after applying a focus measure, a vector consisting of seven neighboring pixels is populated for each pixel in the image volume. PCA is applied on the AC part of the DCT of each vector in the sequence to transform data into eigenspace. Considering the first feature, as it contains maximum variation, and discarding all others, is employed to compute the depth. Though DCT and PCA are both computationally expensive transformations, the reduction in data elements and algorithm iterations have made the new approach efficient. Experimental results are presented to demonstrate the effectiveness of new method by using three different image sequences.

Radiomics analysis of pulmonary nodules in low‐dose CT for early detection of lung cancer

PURPOSE To develop a radiomics prediction model to improve pulmonary nodule (PN) classification in low-dose CT. To compare the model with the American College of Radiology (ACR) Lung CT Screening Reporting and Data System (Lung-RADS) for early detection of lung cancer. METHODS We examined a set of 72 PNs (31 benign and 41 malignant) from the Lung Image Database Consortium image collection (LIDC-IDRI). One hundred three CT radiomic features were extracted from each PN. Before the model building process, distinctive features were identified using a hierarchical clustering method. We then constructed a prediction model by using a support vector machine (SVM) classifier coupled with a least absolute shrinkage and selection operator (LASSO). A tenfold cross-validation (CV) was repeated ten times (10 × 10-fold CV) to evaluate the accuracy of the SVM-LASSO model. Finally, the best model from the 10 × 10-fold CV was further evaluated using 20 × 5- and 50 × 2-fold CVs. RESULTS The best SVM-LASSO model consisted of only two features: the bounding box anterior-posterior dimension (BB_AP) and the standard deviation of inverse difference moment (SD_IDM). The BB_AP measured the extension of a PN in the anterior-posterior direction and was highly correlated (r = 0.94) with the PN size. The SD_IDM was a texture feature that measured the directional variation of the local homogeneity feature IDM. Univariate analysis showed that both features were statistically significant and discriminative (P = 0.00013 and 0.000038, respectively). PNs with larger BB_AP or smaller SD_IDM were more likely malignant. The 10 × 10-fold CV of the best SVM model using the two features achieved an accuracy of 84.6% and 0.89 AUC. By comparison, Lung-RADS achieved an accuracy of 72.2% and 0.77 AUC using four features (size, type, calcification, and spiculation). The prediction improvement of SVM-LASSO comparing to Lung-RADS was statistically significant (McNemars test P = 0.026). Lung-RADS misclassified 19 cases because it was mainly based on PN size, whereas the SVM-LASSO model correctly classified 10 of these cases by combining a size (BB_AP) feature and a texture (SD_IDM) feature. The performance of the SVM-LASSO model was stable when leaving more patients out with five- and twofold CVs (accuracy 84.1% and 81.6%, respectively). CONCLUSION We developed an SVM-LASSO model to predict malignancy of PNs with two CT radiomic features. We demonstrated that the model achieved an accuracy of 84.6%, which was 12.4% higher than Lung-RADS.

Quantifying local tumor morphological changes with Jacobian map for prediction of pathologic tumor response to chemo-radiotherapy in locally advanced esophageal cancer

We proposed a framework to detect and quantify local tumor morphological changes due to chemo-radiotherapy (CRT) using a Jacobian map and to extract quantitative radiomic features from the Jacobian map to predict the pathologic tumor response in locally advanced esophageal cancer patients. In 20 patients who underwent CRT, a multi-resolution BSpline deformable registration was performed to register the follow-up (post-CRT) CT to the baseline CT image. The Jacobian map (J) was computed as the determinant of the gradient of the deformation vector field. The Jacobian map measured the ratio of local tumor volume change where J  <  1 indicated tumor shrinkage and J  >  1 denoted expansion. The tumor was manually delineated and corresponding anatomical landmarks were generated on the baseline and follow-up images. Intensity, texture and geometry features were then extracted from the Jacobian map of the tumor to quantify tumor morphological changes. The importance of each Jacobian feature in predicting pathologic tumor response was evaluated by both univariate and multivariate analysis. We constructed a multivariate prediction model by using a support vector machine (SVM) classifier coupled with a least absolute shrinkage and selection operator (LASSO) for feature selection. The SVM-LASSO model was evaluated using ten-times repeated 10-fold cross-validation (10  ×  10-fold CV). After registration, the average target registration error was 4.30  ±  1.09 mm (LR:1.63 mm AP:1.59 mm SI:3.05 mm) indicating registration error was within two voxels and close to 4 mm slice thickness. Visually, the Jacobian map showed smoothly-varying local shrinkage and expansion regions in a tumor. Quantitatively, the average median Jacobian was 0.80  ±  0.10 and 1.05  ±  0.15 for responder and non-responder tumors, respectively. These indicated that on average responder tumors had 20% median volume shrinkage while non-responder tumors had 5% median volume expansion. In univariate analysis, the minimum Jacobian (p  =  0.009, AUC  =  0.98) and median Jacobian (p  =  0.004, AUC  =  0.95) were the most significant predictors. The SVM-LASSO model achieved the highest accuracy when these two features were selected (sensitivity  =  94.4%, specificity  =  91.8%, AUC  =  0.94). Novel features extracted from the Jacobian map quantified local tumor morphological changes using only baseline tumor contour without post-treatment tumor segmentation. The SVM-LASSO model using the median Jacobian and minimum Jacobian achieved high accuracy in predicting pathologic tumor response. The Jacobian map showed great potential for longitudinal evaluation of tumor response.

SU-C-207B-06: Comparison of Registration Methods for Modeling Pathologic Response of Esophageal Cancer to Chemoradiation Therapy

PURPOSE To compare linear and deformable registration methods for evaluation of tumor response to Chemoradiation therapy (CRT) in patients with esophageal cancer. METHODS Linear and multi-resolution BSpline deformable registration were performed on Pre-Post-CRT CT/PET images of 20 patients with esophageal cancer. For both registration methods, we registered CT using Mean Square Error (MSE) metric, however to register PET we used transformation obtained using Mutual Information (MI) from the same CT due to being multi-modality. Similarity of Warped-CT/PET was quantitatively evaluated using Normalized Mutual Information and plausibility of DF was assessed using inverse consistency Error. To evaluate tumor response four groups of tumor features were examined: (1) Conventional PET/CT e.g. SUV, diameter (2) Clinical parameters e.g. TNM stage, histology (3)spatial-temporal PET features that describe intensity, texture and geometry of tumor (4)all features combined. Dominant features were identified using 10-fold cross-validation and Support Vector Machine (SVM) was deployed for tumor response prediction while the accuracy was evaluated by ROC Area Under Curve (AUC). RESULTS Average and standard deviation of Normalized mutual information for deformable registration using MSE was 0.2±0.054 and for linear registration was 0.1±0.026, showing higher NMI for deformable registration. Likewise for MI metric, deformable registration had 0.13±0.035 comparing to linear counterpart with 0.12±0.037. Inverse consistency error for deformable registration for MSE metric was 4.65±2.49 and for linear was 1.32±2.3 showing smaller value for linear registration. The same conclusion was obtained for MI in terms of inverse consistency error. AUC for both linear and deformable registration was 1 showing no absolute difference in terms of response evaluation. CONCLUSION Deformable registration showed better NMI comparing to linear registration, however inverse consistency of transformation was lower in linear registration. We do not expect to see significant difference when warping PET images using deformable or linear registration. This work was supported in part by the National Cancer Institute Grants R01CA172638.