Jay A. Baker

Training neural network classifiers for medical decision making: the effects of imbalanced datasets on classification performance.

This study investigates the effect of class imbalance in training data when developing neural network classifiers for computer-aided medical diagnosis. The investigation is performed in the presence of other characteristics that are typical among medical data, namely small training sample size, large number of features, and correlations between features. Two methods of neural network training are explored: classical backpropagation (BP) and particle swarm optimization (PSO) with clinically relevant training criteria. An experimental study is performed using simulated data and the conclusions are further validated on real clinical data for breast cancer diagnosis. The results show that classifier performance deteriorates with even modest class imbalance in the training data. Further, it is shown that BP is generally preferable over PSO for imbalanced training data especially with small data sample and large number of features. Finally, it is shown that there is no clear preference between oversampling and no compensation approach and some guidance is provided regarding a proper selection.

Optimized image acquisition for breast tomosynthesis in projection and reconstruction space

Breast tomosynthesis has been an exciting new development in the field of breast imaging. While the diagnostic improvement via tomosynthesis is notable, the full potential of tomosynthesis has not yet been realized. This may be attributed to the dependency of the diagnostic quality of tomosynthesis on multiple variables, each of which needs to be optimized. Those include dose, number of angular projections, and the total angular span of those projections. In this study, the authors investigated the effects of these acquisition parameters on the overall diagnostic image quality of breast tomosynthesis in both the projection and reconstruction space. Five mastectomy specimens were imaged using a prototype tomosynthesis system. 25 angular projections of each specimen were acquired at 6.2 times typical single-view clinical dose level. Images at lower dose levels were then simulated using a noise modification routine. Each projection image was supplemented with 84 simulated 3 mm 3D lesions embedded at the center of 84 nonoverlapping ROIs. The projection images were then reconstructed using a filtered backprojection algorithm at different combinations of acquisition parameters to investigate which of the many possible combinations maximizes the performance. Performance was evaluated in terms of a Laguerre-Gauss channelized Hotelling observer model-based measure of lesion detectability. The analysis was also performed without reconstruction by combining the model results from projection images using Bayesian decision fusion algorithm. The effect of acquisition parameters on projection images and reconstructed slices were then compared to derive an optimization rule for tomosynthesis. The results indicated that projection images yield comparable but higher performance than reconstructed images. Both modes, however, offered similar trends: Performance improved with an increase in the total acquisition dose level and the angular span. Using a constant dose level and angular span, the performance rolled off beyond a certain number of projections, indicating that simply increasing the number of projections in tomosynthesis may not necessarily improve its performance. The best performance for both projection images and tomosynthesis slices was obtained for 15-17 projections spanning an angular are of approximately 45 degrees--the maximum tested in our study, and for an acquisition dose equal to single-view mammography. The optimization framework developed in this framework is applicable to other reconstruction techniques and other multiprojection systems.

Effect of patient histoy data on the prediction of breast cancer from mammographic findings with artificial neural networks

Rationale and Objectives. The authors evaluated the contribution of medical history data to the prediction of breast cancer with artificial neural network (ANN) models based on mammographic findings. Materials and Methods. Three ANNs were developed: The first used 10 Breast Imaging Reporting and Data System (BI-RADS) variables; the second, the BI-RADS variables plus patient age; the third, the BI-RADS variables, patient age, and seven other history variables, for a total of 18 inputs. Performance of the ANNs and the original radiologists impression were evaluated with five metrics: receiver operating characteristic area index (Az); specificity at given sensitivities of 100%, 98%, and 95%; and positive predictive value. Results. All three ANNs consistently outperformed the radiologists impression over all five performance metrics. The patient-age variable was particularly valuable. Adding the age variable to the basic ANN model, which used only the BI-RADS findings, significantly improved Az (P = .028). In fact, replacing all history data with just the age variable resulted in virtually no changes for Az or specificity at 98% sensitivity (P = .324 and P = .410, respectively). Conclusion. Patient age was an important variable for the prediction of breast cancer from mammographic findings with the ANNs. For this data set, all history data could be replaced with age alone.

A mathematical model platform for optimizing a multiprojection breast imaging system

Multiprojection imaging is a technique in which a plurality of digital radiographic images of the same patient are acquired within a short interval of time from slightly different angles. Information from each image is combined to determine the final diagnosis. Projection data are either reconstructed into slices as in the case of tomosynthesis or analyzed directly as in the case of multiprojection correlation imaging technique, thereby avoiding reconstruction artifacts. In this study, the authors investigated the optimum geometry of acquisitions of a multiprojection breast correlation imaging system in terms of the number of projections and their total angular span that yield maximum performance in a task that models clinical decision. Twenty-five angular projections of each breast from 82 human subjects in our breast tomosynthesis database were each supplemented with a simulated 3 mm mass. An approach based on Laguerre-Gauss channelized Hotelling observer was developed to assess the detectability of the mass in terms of receiver operating characteristic (ROC) curves. Two methodologies were developed to integrate results from individual projections into one combined ROC curve as the overall figure of merit. To optimize the acquisition geometry, different components of acquisitions were changed to investigate which one of the many possible configurations maximized the area under the combined ROC curve. Optimization was investigated under two acquisition dose conditions corresponding to a fixed total dose delivered to the patient and a variable dose condition, based on the number of projections used. In either case, the detectability was dependent on the number of projections used, the total angular span of those projections, and the acquisition dose level. In the first case, the detectability approximately followed a bell curve as a function of the number of projections with the maximum between 8 and 16 projections spanning angular arcs of about 23 degrees-45 degrees, respectively. In the second case, the detectability increased with the number of projections approaching an asymptote at 11-17 projections for an angular span of about 45 degrees. These results indicate the inherent information content of the multi-projection image data reflecting the relative role of quantum and anatomical noise in multiprojection breast imaging. The optimization scheme presented here may be applied to any multiprojection imaging modalities and may be extended by including reconstruction in the case of digital breast tomosynthesis and breast computed tomography.

MRI-guided vacuum-assisted breast biopsy with a handheld portable biopsy system.

OBJECTIVE The purpose of this study was to evaluate a compact portable 10-gauge handheld battery-operated vacuum-assisted biopsy system for MRI-guided breast biopsy. CONCLUSION The compact portable battery-operated biopsy system can be used successfully for MRI-guided core breast biopsy and is an alternative to current systems.

Automated breast mass detection in 3D reconstructed tomosynthesis volumes: A featureless approach

The purpose of this study was to propose and implement a computer aided detection (CADe) tool for breast tomosynthesis. This task was accomplished in two stages-a highly sensitive mass detector followed by a false positive (FP) reduction stage. Breast tomosynthesis data from 100 human subject cases were used, of which 25 subjects had one or more mass lesions and the rest were normal. For stage 1, filter parameters were optimized via a grid search. The CADe identified suspicious locations were reconstructed to yield 3D CADe volumes of interest. The first stage yielded a maximum sensitivity of 93% with 7.7 FPs/breast volume. Unlike traditional CADe algorithms in which the second stage FP reduction is done via feature extraction and analysis, instead information theory principles were used with mutual information as a similarity metric. Three schemes were proposed, all using leave-one-case-out cross validation sampling. The three schemes, A, B, and C, differed in the composition of their knowledge base of regions of interest (ROIs). Scheme As knowledge base was comprised of all the mass and FP ROIs generated by the first stage of the algorithm. Scheme B had a knowledge base that contained information from mass ROIs and randomly extracted normal ROIs. Scheme C had information from three sources of information-masses, FPs, and normal ROIs. Also, performance was assessed as a function of the composition of the knowledge base in terms of the number of FP or normal ROIs needed by the system to reach optimal performance. The results indicated that the knowledge base needed no more than 20 times as many FPs and 30 times as many normal ROIs as masses to attain maximal performance. The best overall system performance was 85% sensitivity with 2.4 FPs per breast volume for scheme A, 3.6 FPs per breast volume for scheme B, and 3 FPs per breast volume for scheme C.

Does image quality matter? Impact of resolution and noise on mammographic task performance

The purpose of this study was to examine the effects of different resolution and noise levels on task performance in digital mammography. This study created an image set with images at three different resolution levels, corresponding to three digital display devices, and three different noise levels, with noise magnitudes similar to full clinical dose, half clinical dose, and quarter clinical dose. The images were read by five experienced breast imaging radiologists. The data were then analyzed to compute two accuracy statistics (overall classification accuracy and lesion detection accuracy) and performance at four diagnostic tasks (detection of microcalcifications, benign masses, malignant masses, and discrimination of benign and malignant masses). Human observer results showed decreasing display resolution had little effect on overall classification accuracy and individual diagnostic task performance, but increasing noise caused overall classification accuracy to decrease by a statistically significant 21% as the breast dose went to one quarter of its normal clinical value. The noise effects were most prominent for the tasks of microcalcification detection and mass discrimination. When the noise changed from full clinical dose to quarter clinical dose, the microcalcification detection performance fell from 89% to 67% and the mass discrimination performance decreased from 93% to 79%, while malignant mass detection performance remained relatively constant with values of 88% and 84%, respectively. As a secondary aim, the image set was also analyzed by two observer models to examine whether their performance was similar to humans. Observer models differed from human observers and each other in their sensitivity to resolution degradation and noise. The primary conclusions of this study suggest that quantum noise appears to be the dominant image quality factor in digital mammography, affecting radiologist performance much more profoundly than display resolution.

Is Surgical Excision of Core Biopsy Proven Benign Papillomas of the Breast Necessary

RATIONALE AND OBJECTIVES The aim of this study was to determine if core biopsy-proven benign papillomas of the breast need to be surgically excised. MATERIALS AND METHODS Mammographic and pathologic database review from January 1994 to January 2004 revealed 178 papillary lesions diagnosed by core biopsy in 176 women (mean age, 59 years). All lesions had >or=24 months of imaging follow-up (n = 75) or surgical correlation (n = 103). Details regarding core biopsy technique, lesion appearance, pathologic results, imaging-histopathologic concordance, and follow-up imaging were recorded. Core and surgical pathologic results were correlated. RESULTS Of the 178 papillary lesions diagnosed at core needle biopsy, 120 (67%) were initially diagnosed as benign without atypia. The core biopsy diagnoses of benignity were confirmed for all 120 lesions by either surgical excision (n = 45) or stability after >or=2 years of imaging follow-up (n = 75). Of the remaining 58 papillary lesions, 50 were found to be atypical at core needle biopsy; 15 of those 50 (29%) were upgraded to malignancies at surgical excision. Eight of the 178 lesions (5%) were initially diagnosed as malignant papillary lesions at core needle biopsy. Seven of these eight (88%) were confirmed malignant at excision. None of the surgically proven cancers was diagnosed as benign at core biopsy. CONCLUSIONS Close imaging follow-up rather than excision of core biopsy-proven benign papillomas was adequate given careful imaging-histopathologic correlation and excision of all atypical and discordant lesions. Individual centers should evaluate their own data and tailor their practices accordingly.

Can Breast Cancer Molecular Subtype Help to Select Patients for Preoperative MR Imaging

PURPOSE To assess whether breast cancer molecular subtype classified by surrogate markers can be used to predict the extent of clinically relevant disease with preoperative breast magnetic resonance (MR) imaging. MATERIALS AND METHODS In this HIPAA-compliant, institutional review board-approved study, informed consent was waived. Preoperative breast MR imaging reports from 441 patients were reviewed for multicentric and/or multifocal disease, lymph node involvement, skin and/or nipple invasion, chest wall and/or pectoralis muscle invasion, or contralateral disease. Pathologic reports were reviewed to confirm the MR imaging findings and for hormone receptors (estrogen and progesterone subtypes), human epidermal growth factor receptor type 2 (HER2 subtype), tumor size, and tumor grade. Surrogates were used to categorize tumors by molecular subtype: hormone receptor positive and HER2 negative (luminal A subtype); hormone receptor positive and HER2 positive (luminal B subtype); hormone receptor negative and HER2 positive (HER2 subtype); hormone receptor negative and HER2 negative (basal subtype). All patients included in the study had a histologic correlation with MR imaging findings or they were excluded. χ(2) analysis was used to compare differences between subtypes, with multivariate logistic regression analysis used to assess for variable independence. RESULTS Identified were 289 (65.5%) luminal A, 45 (10.2%) luminal B, 26 (5.9%) HER2, and 81 (18.4%) basal subtypes. Among subtypes, significant differences were found in the frequency of multicentric and/or multifocal disease (luminal A, 27.3% [79 of 289]; luminal B, 53.3% [24 of 45]; HER2, 65.4% [17 of 26]; basal, 27.2% [22 of 81]; P < .001) and lymph node involvement (luminal A, 17.3% [50 of 289]; luminal B, 35.6% [26 of 45]; HER2, 34.6% [nine of 26]; basal 24.7% [20 of 81]; P = .014). Multivariate analysis showed that molecular subtype was independently predictive of multifocal and/or multicentric disease. CONCLUSION Preoperative breast MR imaging is significantly more likely to help detect multifocal and/or multicentric disease and lymph node involvement in luminal B and HER2 molecular subtype breast cancers. Molecular subtype may help to select patients for preoperative breast MR imaging.

Preoperative Single-Fraction Partial Breast Radiation Therapy: A Novel Phase 1, Dose-Escalation Protocol With Radiation Response Biomarkers.

PURPOSE Women with biologically favorable early-stage breast cancer are increasingly treated with accelerated partial breast radiation (PBI). However, treatment-related morbidities have been linked to the large postoperative treatment volumes required for external beam PBI. Relative to external beam delivery, alternative PBI techniques require equipment that is not universally available. To address these issues, we designed a phase 1 trial utilizing widely available technology to 1) evaluate the safety of a single radiation treatment delivered preoperatively to the small-volume, intact breast tumor and 2) identify imaging and genomic markers of radiation response. METHODS AND MATERIALS Women aged ≥55 years with clinically node-negative, estrogen receptor-positive, and/or progesterone receptor-positive HER2-, T1 invasive carcinomas, or low- to intermediate-grade in situ disease ≤2 cm were enrolled (n=32). Intensity modulated radiation therapy was used to deliver 15 Gy (n=8), 18 Gy (n=8), or 21 Gy (n=16) to the tumor with a 1.5-cm margin. Lumpectomy was performed within 10 days. Paired pre- and postradiation magnetic resonance images and patient tumor samples were analyzed. RESULTS No dose-limiting toxicity was observed. At a median follow-up of 23 months, there have been no recurrences. Physician-rated cosmetic outcomes were good/excellent, and chronic toxicities were grade 1 to 2 (fibrosis, hyperpigmentation) in patients receiving preoperative radiation only. Evidence of dose-dependent changes in vascular permeability, cell density, and expression of genes regulating immunity and cell death were seen in response to radiation. CONCLUSIONS Preoperative single-dose radiation therapy to intact breast tumors is well tolerated. Radiation response is marked by early indicators of cell death in this biologically favorable patient cohort. This study represents a first step toward a novel partial breast radiation approach. Preoperative radiation should be tested in future clinical trials because it has the potential to challenge the current treatment paradigm and provide a path forward to identify radiation response biomarkers.