Assistant Professor, University of Southern California
Our projects often involve developing of image-processing pipelines, feature analysis toolkits, including but not limited to shape and texture analysis, tumor identification and classification based on tumor type and grade of malignancy, etc., across various diseases and using a variety of imaging and non-imaging data. The resultant tumor behavior models relating imaging features to tumor behavior is potentially of great value as a research tool and a clinical decision support tool, to improve individualized treatment selection and aid advanced treatment monitoring. Results from our projects feed into our RADIOMICS platform -the high throughput extraction of tumor features using standard-of-care imaging platform, where imaging features can be combined with clinical, laboratory, genomic, and epigenetic data to improve identification of diagnostic and prognostic features. A comprehensive capture of tumor heterogeneity via the capture of its various phenotypes using standard-of-care imaging will aid in advancement of precision medicine. For the RSNA conference , we will present the improvements of our Radiomics platform by adding wavelet transforms. Our preliminary results show a significant difference (P<0.01) in the texture profile of renal benign masses versus malignant masses based on wavelet analysis.
Abstract: Finite element (FE) models of long bones constructed from computed-tomography (CT) data are emerging as an invaluable tool in the field of bone biomechanics. However, the performance of such FE models is highly dependent on the accurate capture of geometry and appropriate assignment of material properties. In this study, a combined numerical-experimental study is performed comparing FE-predicted surface strains with strain-gauge measurements. Thirty-six major, cadaveric, long bones (humerus, radius, femur and tibia), which cover a wide range of bone sizes, were tested under three-point bending and torsion. The FE models were constructed from trans-axial volumetric CT scans, and the segmented bone images were corrected for partial-volume effects. The material properties (Young's modulus for cortex, density-modulus relationship for trabecular bone and Poisson's ratio) were calibrated by minimizing the error between experiments and simulations among all bones. The R(2) values of the measured strains versus load under three-point bending and torsion were 0.96-0.99 and 0.61-0.99, respectively, for all bones in our dataset. The errors of the calculated FE strains in comparison to those measured using strain gauges in the mechanical tests ranged from -6% to 7% under bending and from -37% to 19% under torsion. The observation of comparatively low errors and high correlations between the FE-predicted strains and the experimental strains, across the various types of bones and loading conditions (bending and torsion), validates our approach to bone segmentation and our choice of material properties.
Pub.: 04 Feb '11, Pinned: 19 Jun '17
Abstract: The aim of this study was to develop quantitative computed-tomography (QCT)-based bone-strength indicators that highly correlate with finite-element (FE)-based strength. Transaxial QCT scans were obtained from 36 major, cadaveric, long bones (humerus, radius, femur and tibia) from 4 females and 2 males, 53 to 86 years old. These images were used to construct the FE models and to develop the QCT-based bone strength indicators under every-day, simplified loading conditions. We have evaluated the performance of area-weighted (AW), density-weighted (DW) and modulus-weighted (MW) rigidity measures as well as popular strength indicators like section modulus (Z) and stress-strain index (SSI). We have also developed a novel strength metric, the centroid deviation, which analyzes the spatial distribution of the centroids along the length of the bone. The correlation results show that the MW polar moment of inertia and the MW moment of inertia are the two top-performers for all bones and loading conditions (average r>0.89). The MW centroid deviations correlated highly with the estimated load to fracture for all bones under compression (r>0.83), except for the humerus (r=0.67). Consistently DW or MW rigidity measures produced a statistically significant improvement in capturing bone strength compared to AW rigidity measures. As expected, MW rigidity measures showed a higher correlation with the FE-based fracture load than the DW rigidity measures; however, the improvement was not statistically significant. Through this study we present a short-list of useful QCT-based strength parameters that correlate well with FE-based fracture load. Although a few parameters perform reasonably well across most bones and loading conditions, a judicious assessment of bone strength should include multiple parameters evaluated at multiple critical locations in the long bones, with attention to the type of loading and bone type.
Pub.: 01 Nov '11, Pinned: 19 Jun '17
Abstract: The purpose of this study was to compare whole-lesion (WL) enhancement parameters to single region of interest (ROI)-based enhancement in discriminating clear cell renal cell carcinoma (ccRCC) from renal oncocytoma.In this IRB-approved retrospective study, the surgical database was queried to derive a cohort of 94 postnephrectomy patients with ccRCC or oncocytoma (68 ccRCC, 26 oncocytoma), who underwent preoperative multiphase contrast-enhanced computed tomography (CECT) between June 2009 and August 2013. CT acquisitions were transferred to a three-dimensional workstation, and WL ROIs were manually segmented. WL enhancement and histogram distribution parameters skewness, kurtosis, standard deviation (SD), and interquartile range (IQR) were calculated. WL enhancement parameters were compared to single ROI-based enhancement using receiver operating characteristic (ROC) analysis.Oncocytoma had significantly higher WL enhancement than ccRCC in nephrographic (mean, p = 0.02; median, p = 0.03) and excretory phases (mean, p = 0.03; median p < 0.01). ccRCC had significantly higher kurtosis than oncocytoma in corticomedullary (p = 0.03) and excretory phases (p < 0.01), and significantly higher SD and IQR than oncocytoma in all postcontrast phases: corticomedullary (SD, p = 0.02; IQR, p < 0.01), nephrographic (SD, p = 0.01; IQR, p = 0.03), and excretory (SD, p < 0.01; IQR, p < 0.01). When compared to single ROI-based enhancement, WL enhancement alone did not demonstrate a statistical advantage in discriminating between ccRCC and oncocytoma (area under ROC curve of 0.78 and 0.72 respectively), but when combined with histogram distribution parameters (area under ROC curve of 0.86), it did demonstrate a slight improvement.Our study suggests that voxel-based WL enhancement parameters provide only a slight improvement over single ROI-based enhancement techniques in differentiating between ccRCC and renal oncocytoma.
Pub.: 07 Sep '16, Pinned: 19 Jun '17