Andrew R. Jamieson

Evaluation of Computer-aided Diagnosis on a Large Clinical Full-field Digital Mammographic Dataset

RATIONALE AND OBJECTIVES To convert and optimize our previously developed computerized analysis methods for use with images from full-field digital mammography (FFDM) for breast mass classification to aid in the diagnosis of breast cancer. MATERIALS AND METHODS An institutional review board approved protocol was obtained, with waiver of consent for retrospective use of mammograms and pathology data. Seven hundred thirty-nine FFDM images, which contained 287 biopsy-proven breast mass lesions, of which 148 lesions were malignant and 139 lesions were benign, were retrospectively collected. Lesion margins were delineated by an expert breast radiologist and were used as the truth for lesion-segmentation evaluation. Our computerized image analysis method consisted of several steps: 1) identified lesions were automatically extracted from the parenchymal background using computerized segmentation methods; 2) a set of image characteristics (mathematic descriptors) were automatically extracted from image data of the lesions and surrounding tissues; and 3) selected features were merged into an estimate of the probability of malignancy using a Bayesian artificial neural network classifier. Performance of the analyses was evaluated at various stages of the conversion using receiver-operating characteristic analysis. RESULTS An area under the curve value of 0.81 was obtained in the task of distinguishing between malignant and benign mass lesions in a round-robin by case evaluation on the entire FFDM dataset. We failed to show a statistically significant difference (P = .83) compared to results from our previous study in which the computerized classification was performed on digitized screen-film mammograms. CONCLUSIONS Our computerized analysis methods developed on digitized screen-film mammography can be converted for use with FFDM. Results show that the computerized analysis methods for the diagnosis of breast mass lesions on FFDM are promising, and can potentially be used to aid clinicians in the diagnostic interpretation of FFDM.

Breast image feature learning with adaptive deconvolutional networks

Feature extraction is a critical component of medical image analysis. Many computer-aided diagnosis approaches employ hand-designed, heuristic lesion extracted features. An alternative approach is to learn features directly from images. In this preliminary study, we explored the use of Adaptive Deconvolutional Networks (ADN) for learning high-level features in diagnostic breast mass lesion images with potential application to computer-aided diagnosis (CADx) and content-based image retrieval (CBIR). ADNs (Zeiler, et. al., 2011), are recently-proposed unsupervised, generative hierarchical models that decompose images via convolution sparse coding and max pooling. We trained the ADNs to learn multiple layers of representation for two breast image data sets on two different modalities (739 full field digital mammography (FFDM) and 2393 ultrasound images). Feature map calculations were accelerated by use of GPUs. Following Zeiler et. al., we applied the Spatial Pyramid Matching (SPM) kernel (Lazebnik, et. al., 2006) on the inferred feature maps and combined this with a linear support vector machine (SVM) classifier for the task of binary classification between cancer and non-cancer breast mass lesions. Non-linear, local structure preserving dimension reduction, Elastic Embedding (Carreira-Perpiñán, 2010), was then used to visualize the SPM kernel output in 2D and qualitatively inspect image relationships learned. Performance was found to be competitive with current CADx schemes that use human-designed features, e.g., achieving a 0.632+ bootstrap AUC (by case) of 0.83 [0.78, 0.89] for an ultrasound image set (1125 cases).

Comparison of computerized image analyses for digitized screen-film mammograms and full-field digital mammography images

We have developed computerized methods for the analysis of mammographic lesions in order to aid in the diagnosis of breast cancer. Our automatic methods include the extraction of the lesion from the breast parenchyma, the characterization of the lesion features in terms of mathematical descriptors, and the merging of these lesion features into an estimate of the probability of malignancy. Our initial development was performed on digitized screen film mammograms. We report our progress here in converting our methods for use with images from full-field digital mammography (FFDM). It is apparent from our initial comparisons on CAD for SFMD and FFDM that the overall concepts and image analysis techniques are similar, however reoptimization for a particular lesion segmentation or a particular mammographic imaging system are warranted.

TU‐D‐330A‐03: Comparison of Image Segmentation Algorithms On Digitized Mammograms and FFDM Images for CAD

Purpose: To investigate lesion tumor segmentation methods for both digitized screen‐film mammograms (DSF) and full‐field digital mammography (FFDM). Method and Materials: Breast lesion segmentation methods are important in the overall image analysis for computer‐aided diagnosis. Our initial development was performed on DSF, and we are currently evaluating our methods for use with FFDM. Three of our lesion segmentation methods were investigated using a database of 84 DFM and 287 FFDM cases including malignant and benign lesions. A region growing method utilizes the size and shape of the evolving lesion contour to determine the lesion margin. A radial gradient index (RGI) segmentation method uses a Gaussian constraint function to suppress the influence of distant pixels. Then for a series of contours returned by grey level thresholding, the contour with maximum RGI is chosen as the one that best delineates the lesion. A two‐stage, region‐based active contour method minimizes an energy function based on the homogeneities inside and outside of the evolving contour. The minimization algorithm solves the Euler‐Lagrange equation describing the contour evolution. Prior to the application of the active contour model, RGI segmentation is applied to delineate an initial contour closer to the lesion margin and estimate the effective background. The methods were compared to radiologist‐delineated margins on both DSF and FFDM images using an area similarity metric. Results: At an overlap threshold of 0.3, the region growing, RGI, and two‐stage methods correctly segmented 84%, 86% and 94% of the digitized screen‐film lesions, and 81%, 83% and 88% of lesions on FFDM, respectively. Conclusion: Our results indicate that the two‐stage method yields improved segmentation for both DSF and FFDM, and also that methods developed with DSF can be efficiently converted for use with FFDM. Conflict of Interest: MLG is a shareholder in R2 Technology, Inc.

WE‐D‐304A‐02: Exploring Non‐Linear Feature Space Dimension Reduction and Data Representation in Breast CADx

Purpose: Preliminary study to explore potential of recently developed non‐linear dimension reduction and data representation techniques as applied to breast lesion CADx generated feature spaces. Method and Materials: Two new methods were explored: Laplacian eignenmaps of (Belkin and Niyogi) and t‐distributed stochastic neighbor embedding (t‐SNE) (van der Maaten and Hinton). The properties of these methods were evaluated in the context of malignancy classification performance as well as visual inspection of the sparseness for two and three dimensional mapping representations. The robustness of the proposed techniques were tested against four separate imaging modality feature databases, including 2956 ultrasound(US), 735 full‐field digital mammography (FFDM), 850 screen‐film mammography (SFM), and 356 DCE‐MRI biopsy proven breast mass lesions. Using the reduced dimension mapped feature output as input, as opposed to feature selection, two classifiers, linear and non‐linear, were tested: Markov Chain Monte Carlo based Bayesian artificial neural network (MCMC‐BANN) and linear discriminate analysis (LDA). To evaluate performance, the AUC was estimated for each classifier using ROC analysis and 0.632+ bootstrap validation on over 500 samples. In addition, performance was compared to a non‐linear Automatic Relevance Determination (ARD), linear step‐wise feature selection method, as well as a linear reduction Principle Component Analysis based method. Results: The new methods were found to match or exceed performance of current state‐of‐the‐art breast lesion CADx algorithms for feature selection and classification across all modalities. Additionally, the new techniques possess the added benefit of naturally delivering sparse lower dimensional representations for visual interpretation, thus, revealing intricate data structure of the feature space. Conclusion: These methods maintain predictive power while also preserving both local and global structural information present in the original high dimensional feature space for breast lesion feature data used in CADx.Conflict of Interest: M. Giger is a stockholder and receives royalties from Hologic.

SU‐GG‐I‐04: Grid‐Computing for Optimization of CAD

Purpose: Pilot study to utilize large scale parallel grid computing to harness the nationwide cluster infrastructure for optimization of medical image analysis parameters. Method and Materials: A previously developed CAD scheme for mass lesions in mammography was ported onto the grid computing environment by wrapping the algorithm code with the virtual data language (VDL). The CAD scheme was then configured into a parallelizable workflow by the grid‐software. The workflows were executed using two test clusters (in Santa Monica, CA and Chicago, IL) consisting of over 220 dual‐CPU nodes combined. Using the grid‐environment workflow, parameter sweeps were conducted for lesion segmentation settings based on radial‐gradient‐index (RGI) methods. Specifically, the Gaussian width (GW) used in initially filtering lesion images for segmentation was varied by increments of 1 mm from 1 to 60 mm. For each GW sweep the entire 850 biopsy‐proven mass lesion database (411 benign, 439 malignant) was analyzed. In each, 29 different mathematical descriptor features were calculated, followed by feature selection and merging with linear discriminate analysis. Diagnostic performance was estimated by ROC analysis by calculating AUC (from PROPROC) values based on both individual features alone, and merged. For merged classifiers, AUC values were found using round‐robin case‐by‐case removal and replacement. Results: Computation jobs requiring over 30 CPU hours on a single lab computer were completed in approximately 35 minutes in this preliminary study. Merged AUC values increased from 0.50 (std.err.=0.018) at GW of 1mm with, to 0.81 (std.err.=0.015) at 10mm GW, with relative plateaus across the rest of the parameter space to 60mm. Conclusion: The parameter space sweep in GW identified trends in individual feature performance as well as merged results. Large scale, computationally intensive image analysis can be carried out in a timely fashion, feasible for expedited experimental discovery, as well as for more thorough future statistical analysis.