Christopher B. Kendall

Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning?

Training a deep convolutional neural network (CNN) from scratch is difficult because it requires a large amount of labeled training data and a great deal of expertise to ensure proper convergence. A promising alternative is to fine-tune a CNN that has been pre-trained using, for instance, a large set of labeled natural images. However, the substantial differences between natural and medical images may advise against such knowledge transfer. In this paper, we seek to answer the following central question in the context of medical image analysis: Can the use of pre-trained deep CNNs with sufficient fine-tuning eliminate the need for training a deep CNN from scratch? To address this question, we considered four distinct medical imaging applications in three specialties (radiology, cardiology, and gastroenterology) involving classification, detection, and segmentation from three different imaging modalities, and investigated how the performance of deep CNNs trained from scratch compared with the pre-trained CNNs fine-tuned in a layer-wise manner. Our experiments consistently demonstrated that 1) the use of a pre-trained CNN with adequate fine-tuning outperformed or, in the worst case, performed as well as a CNN trained from scratch; 2) fine-tuned CNNs were more robust to the size of training sets than CNNs trained from scratch; 3) neither shallow tuning nor deep tuning was the optimal choice for a particular application; and 4) our layer-wise fine-tuning scheme could offer a practical way to reach the best performance for the application at hand based on the amount of available data.Training a deep convolutional neural network (CNN) from scratch is difficult because it requires a large amount of labeled training data and a great deal of expertise to ensure proper convergence. A promising alternative is to fine-tune a CNN that has been pre-trained using, for instance, a large set of labeled natural images. However, the substantial differences between natural and medical images may advise against such knowledge transfer. In this paper, we seek to answer the following central question in the context of medical image analysis: Can the use of pre-trained deep CNNs with sufficient fine-tuning eliminate the need for training a deep CNN from scratch? To address this question, we considered four distinct medical imaging applications in three specialties (radiology, cardiology, and gastroenterology) involving classification, detection, and segmentation from three different imaging modalities, and investigated how the performance of deep CNNs trained from scratch compared with the pre-trained CNNs fine-tuned in a layer-wise manner. Our experiments consistently demonstrated that 1) the use of a pre-trained CNN with adequate fine-tuning outperformed or, in the worst case, performed as well as a CNN trained from scratch; 2) fine-tuned CNNs were more robust to the size of training sets than CNNs trained from scratch; 3) neither shallow tuning nor deep tuning was the optimal choice for a particular application; and 4) our layer-wise fine-tuning scheme could offer a practical way to reach the best performance for the application at hand based on the amount of available data.

Parachute Mitral Valve in Adults—A Systematic Overview

Parachute mitral valve (PMV) is a rare congenital anomaly of the mitral valve apparatus seen in infants and young children. In most instances PMV is associated with other congenital anomalies of the heart, in particular obstructive lesions of the mitral inflow (mitral valve ring) and left ventricular outflow tract (subaortic stenosis), and coarctation of aorta and is referred to as Shones complex or Shones anomaly. PMV may also occur as an isolated lesion or in association with other congenital cardiac anomalies. Not much is known about PMV in adults as an isolated anomaly or in association with other congenital cardiac anomalies. We reviewed the literature to identify cases of PMV (isolated or associated with other lesions) in adults, to address prevalence, clinical presentation, diagnosis, treatment, and outcome of such patients. (Echocardiography 2010;27:581‐586)

Incidence of subclinical atherosclerosis as a marker of cardiovascular risk in retired professional football players.

The purpose of this study was to evaluate subclinical atherosclerosis in retired professional football players. Two hundred one healthy former professional football players (mean age 50.8 years; mean body mass index 31.5 kg/m(2)) were screened for the prevalence of cardiovascular risk factors, metabolic syndrome, and subclinical atherosclerosis by carotid artery ultrasound and compared with a cohort of men of similar body mass index referred for the assessment of subclinical atherosclerosis by carotid ultrasound. The prevalence of carotid artery plaque in the players was not significantly different from that of the body mass index-matched patients (33.3% vs 29.3%, p = 0.45). For the 2 groups, the prevalence of carotid artery plaque was >3 times higher than that reported in general population studies of patients with the same age range, gender, and exclusions. Metabolic syndrome prevalence was higher in linemen than in nonlinemen (45.8% vs 22.5%, p = 0.001), but there was no statistical difference in plaque presence between linemen and nonlinemen (27.1% vs 35.9%, p = 0.23). In conclusion, despite their elite athletic histories, former professional football players have a similar prevalence of advanced subclinical atherosclerosis as a clinically referred population of overweight and obese men.

Papillary Fibroelastoma of the Pulmonary Valve – A Systematic Review

The pulmonary valve is the least affected site for valvular papillary fibroelastoma. With increasing use of routine echocardiography and other modalities of imaging, pulmonary valve papillary fibroelastomas (PVPFE) are being recognized more frequently. PVPFE is more often an incidental diagnosis and symptomatic patients usually present with shortness of breath. Embolic phenomena and right ventricular outflow tract obstruction are the most serious complications of PVPFE. Since PVPFE is rare, the purpose of this systematic review is to address demographic characteristics, the clinical presentation, management, and outcome of this benign tumor of the pulmonary valve.

Left Ventricular Assist Devices: Physiologic Assessment using Echocardiography for Management and Optimization

Left ventricular assist devices (LVAD) are being deployed increasingly in patients with severe left ventricular dysfunction and medically refractory congestive heart failure of any etiology. The United States Food and Drug Administration (FDA) recently approved the use of the Thoratec Heartmate II (Thoratec Corporation, Pleasanton, CA, USA) for outpatient use. Echocardiography is fundamental during each stage of patient management, pre-LVAD placement, during LVAD placement, for postoperative LVAD optimization and long-term follow-up. We present a pragmatic and systematic echocardiographic approach that serves as a guide for the management of left ventricular assist devices.

Automating Carotid Intima-Media Thickness Video Interpretation with Convolutional Neural Networks

Cardiovascular disease (CVD) is the leading cause of mortality yet largely preventable, but the key to prevention is to identify at-risk individuals before adverse events. For predicting individual CVD risk, carotid intima-media thickness (CIMT), a noninvasive ultrasound method, has proven to be valuable, offering several advantages over CT coronary artery calcium score. However, each CIMT examination includes several ultrasound videos, and interpreting each of these CIMT videos involves three operations: (1) select three end-diastolic ultrasound frames (EUF) in the video, (2) localize a region of interest (ROI) in each selected frame, and (3) trace the lumen-intima interface and the media-adventitia interface in each ROI to measure CIMT. These operations are tedious, laborious, and time consuming, a serious limitation that hinders the widespread utilization of CIMT in clinical practice. To overcome this limitation, this paper presents a new system to automate CIMT video interpretation. Our extensive experiments demonstrate that the suggested system performs reliably. The reliable performance is attributable to our unified framework based on convolutional neural networks (CNNs) coupled with our informative image representation and effective post-processing of the CNN outputs, which are uniquely designed for each of the above three operations.

ECG-based frame selection and curvature-based ROI detection for measuring carotid intima-media thickness

Carotid intima-media thickness (CIMT) has proven to be sensitive for predicting individual risk of cardiovascular diseases (CVD). The CIMT is measured based on region of interest (ROIs) in end-diastolic ultrasound frames (EUFs). To interpret CIMT videos, in the current practice, the EUFs and ROIs must be manually selected, a process that is tedious and time consuming. To reduce CIMT interpretation time, this paper presents a novel method for automatically selecting EUFs and determining ROIs in ultrasound videos. The EUFs are selected based on the QRS complex of the electrocardiogram (ECG) signal associated with the ultrasound video, and the ROI is detected based on image intensity and curvature of the carotid artery bulb. Once a EUF is selected and its corresponding ROI is determined, our system measures CIMT using the snake algorithm extended with hard constraints [1,6-7] by computing the average thickness and maximum thickness, calculating the vascular age, and generating a patient’s report. In this study, we utilize 23 subjects. Each subject has 4 videos, and 3 EUFs are selected in each video, resulting in a total of 272 ROIs. By comparing with the reference provided by an expert for both frame selection and ROI detection, we achieve 92.96% sensitivity and 97.62% specificity for EUF selection, and 81.25% accuracy in ROI detection.

Carotid Intimal–Medial Thickness in Active Professional American Football Players Aged 23 to 35 Years

Risk of cardiovascular disease and death in retired professional American football players may be higher than that in the general population. Previously published data have demonstrated that American football players have less glucose intolerance, less smoking, similar lipid profiles, and higher blood pressure despite a much larger body compared to the general population, although the presence of subclinical atherosclerosis in these subjects has not been evaluated. This study compared the prevalence of subclinical atherosclerosis in active professional American football players to that in age-, gender-, and race-matched controls derived from the Bogalusa Heart Study. Carotid intimal-medial thickness (CIMT) was used as an indicator of subclinical atherosclerosis in 75 active American football players (23 to 35 years old, 31 white, 44 African-American) as measured by B-mode ultrasonography at Mayo Clinic, Scottsdale, Arizona, on September 13 and 14, 2009. CIMT measurements of 75 athletes were compared to those of 518 matched controls who had CIMT determinations in 1995 and 1996. Two-group t tests determined population similarities between groups. In a generalized linear model, players (overall and by race) had lower CIMT values than controls after age and race adjustment (p <0.001 for all comparisons). Nonlinemen and linemen had lower CIMT values than controls (p < 0.001 and p = 0.004, respectively). In conclusion, active professional American football players, regardless of position, had mean CIMT values similar to or lower than those in a matched general population cohort, suggesting that if the prevalence of subclinical atherosclerosis is increased in retired professional American football players, this occurs after retirement.

A user friendly system for ultrasound carotid intima-media thickness image interpretation

Assessment of Carotid Intima-Media Thickness (CIMT) by B-mode ultrasound is a technically mature and reproducible technology. Given the high morbidity, mortality and the large societal burden associated with CV diseases, as a safe yet inexpensive tool, CIMT is increasingly utilized for cardiovascular (CV) risk stratification. However, CIMT requires a precise measure of the thickness of the intima and media layers of the carotid artery that can be tedious, time consuming, and demand specialized expertise and experience. To this end, we have developed a highly user-friendly system for semiautomatic CIMT image interpretation. Our contribution is the application of active contour models (snake models) with hard constraints, leading to an accurate, adaptive and user-friendly border detection algorithm. A comparison study with the CIMT measurement software in Siemens Syngo® Arterial Health Package shows that our system gives a small bias in mean (0.049 ±0.051mm) and maximum (0.010 ± 0.083 mm) CIMT measures and offers a higher reproducibility (average correlation coefficients were 0.948 and 0.844 in mean and maximum CIMT respectively (P <0.001)). This superior performance is attributed to our novel interface design for hard constraints in the snake models.

Risk Stratification for Cardiovascular Disease in Women in the Primary Care Setting

BACKGROUND Traditional risk assessment tools classify the majority of middle-aged women at low risk despite cardiovascular (CV) diseases affecting >50% of women and remaining the leading cause of death. Ultrasound-determined carotid intima-media thickness (CIMT) and/or computed tomographic coronary artery calcium score (CACS) quantify subclinical atherosclerosis and add incremental prognostic value. The aim of this study was to assess the utility of CIMT and CACS to detect subclinical atherosclerosis in younger women. METHODS Asymptomatic women aged 50 to 65 years with at least one CV risk factor and low Framingham risk scores were identified prospectively at primary care and cardiology clinics. Mean intimal thickness, plaque on CIMT, and Agatston calcium score for CACS were obtained. RESULTS Of 86 women (mean age, 58 ± 4.6 years; mean Framingham risk score, 1.9 ± 1.2; mean low-density lipoprotein cholesterol level, 138.9 ± 37.0 mg/dL), 53 (62%) had high-risk CIMT (51% plaque, 11% CIMT > 75th percentile). In contrast, three women (3.5%) had CACS > 100, all of whom had plaque by CIMT. Of the 58 women with CACS of 0, 32 (55%) had high-risk CIMT (48% plaque, 7% CIMT > 75th percentile). CONCLUSIONS In patients referred by their physicians for assessment of CV risk, CIMT in asymptomatic middle-aged women with at least one CV risk factor and low risk by the Framingham risk score identified a large number with advanced subclinical atherosclerosis despite low CACS. Our results suggest that CIMT may be a more sensitive method for CV risk assessment than CACS or traditional risk tools in this population. Further studies are needed to determine if earlier detection would be of clinical benefit.