R. Todd Hurst

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.

Takotsubo Cardiomyopathy: A Unique Cardiomyopathy With Variable Ventricular Morphology

Takotsubo cardiomyopathy is an important differential diagnosis of acute coronary syndrome. It is characterized by normal (or near-normal) coronary arteries, regional wall motion abnormalities that extend beyond a single coronary vascular bed, and often, a precipitating stressor. Variants of the classical left ventricular apical ballooning, including mid- or basal left ventricular wall motion abnormalities, are increasingly recognized. Takotsubo cardiomyopathy is not rare, and heightened awareness of this unique cardiomyopathy likely will lead to a higher reported incidence. Diagnosis of takotsubo cardiomyopathy has important implications for clinical management at presentation and afterward. The long-term prognosis is generally favorable; however, a small subset has potentially life-threatening complications during the initial presentation. The pathophysiologic mechanism is unknown, but catecholamine excess likely has a central role.

Noninvasive Measurement of Central Vascular Pressures With Arterial Tonometry: Clinical Revival of the Pulse Pressure Waveform?

The arterial pulse has historically been an essential source of information in the clinical assessment of health. With current sphygmomanometric and oscillometric devices, only the peak and trough of the peripheral arterial pulse waveform are clinically used. Several limitations exist with peripheral blood pressure. First, central aortic pressure is a better predictor of cardiovascular outcome than peripheral pressure. Second, peripherally obtained blood pressure does not accurately reflect central pressure because of pressure amplification. Lastly, antihypertensive medications have differing effects on central pressures despite similar reductions in brachial blood pressure. Applanation tonometry can overcome the limitations of peripheral pressure by determining the shape of the aortic waveform from the radial artery. Waveform analysis not only indicates central systolic and diastolic pressure but also determines the influence of pulse wave reflection on the central pressure waveform. It can serve as a useful adjunct to brachial blood pressure measurements in initiating and monitoring hypertensive treatment, in observing the hemodynamic effects of atherosclerotic risk factors, and in predicting cardiovascular outcomes and events. Radial artery applanation tonometry is a noninvasive, reproducible, and affordable technology that can be used in conjunction with peripherally obtained blood pressure to guide patient management. Keywords for the PubMed search were applanation tonometry, radial artery, central pressure, cardiovascular risk, blood pressure, and arterial pulse. Articles published from January 1, 1995, to July 1, 2009, were included in the review if they measured central pressure using radial artery applanation tonometry.

Increased Incidence of Coronary Atherosclerosis in Type 2 Diabetes Mellitus: Mechanisms and Management

Type 2 diabetes mellitus is a vascular disease. More than 3 out of 4 diabetic patients die of causes related to atherosclerosis, in most cases (75%) because of coronary artery disease (1). Yet, 70% of diabetic persons do not believe they are at serious risk for cardiovascular disease (2). An increasing number of patients have diabetes and its attendant complications, and this trend is predicted to continue. In the United States, the number of diabetic patients is expected to increase from 15 million to 22 million by 2025 (1). This increase correlates strongly with increasing obesity (3) and is reflected in the development of diabetes at an earlier age (4). The younger age at onset portends an increased future prevalence of premature coronary artery disease and resource utilization that will exceed the current

Carotid Intima-Media Thickness and Coronary Artery Calcium Score as Indications of Subclinical Atherosclerosis

100 billion annual expenditure in the United States (5). Type 2 diabetes increases the risk for coronary artery disease by 2 to 4 times in the overall population. Haffner and colleagues (6) found that diabetic patients with no history of coronary artery disease have the same risk for future myocardial infarction as do nondiabetic patients with known disease (Figure). The National Cholesterol Education Program considers diabetes to be a coronary disease equivalent in their lipid guidelines (7). The risk is even greater in women. Diabetes eliminates the usual female advantage in risk for death from coronary artery disease; these patients have a 5-fold to 8-fold higher death rate than do nondiabetic women (8). Although advances in cardiovascular care and revascularization techniques have decreased event rates and mortality in the population overall, these rates are increasing among diabetic patients (9). Figure. KaplanMeier estimates of the probability of death from coronary heart disease in 1059 patients with type 2 diabetes and 1378 nondiabetic patients with and without previous myocardial infarction ( MI ). The purpose of this article is to define the mechanisms for the propensity for atherosclerosis, emphasize the benefits of prevention, and review the optimal management of coronary atherosclerosis in patients with type 2 diabetes. For details of our literature search strategy, see the Appendix. The Metabolic Syndrome The metabolic syndrome, also known as the insulin-resistance syndrome or syndrome X (Table 1), affects 50 to 75 million Americans (10) and often precedes diabetes. Twenty percent of middle-aged persons and 40% of older persons are affected (10). These patients are not frankly diabetic but still have substantial risk for atherosclerosis, with rates of total and cardiovascular mortality that are double those in the general population (11). Table 1. The National Cholesterol Education Program Guidelines for Diagnosis of the Metabolic Syndrome The same therapies that decrease coronary risk may also be of benefit in preventing the onset of diabetes and its complications. The Diabetes Prevention Program (12) study randomly assigned 3234 patients with elevated fasting and postload plasma glucose levels to receive placebo or metformin or follow a lifestyle modification program. Lifestyle modifications (diet, exercise, and weight loss) decreased the incidence of newly diagnosed diabetes by 58%, and metformin therapy produced a significant decrease of 31% (12). The clinical benefits of treatment with angiotensin-converting enzyme inhibitors (13) and statins (14, 15) in the prevention of cardiac events are well established. These drugs may also delay or prevent onset of diabetes (16, 17). The Diabetes Reduction Assessment with Ramipril and Rosiglitizone Medication study will evaluate whether therapy with an angiotensin-converting enzyme inhibitor, possibly by improving insulin sensitivity, can decrease the incidence of diabetes in persons who have impaired glucose tolerance. Aggressive management of coexisting cardiac risk factors is paramount in the metabolic syndrome. The National Cholesterol Education Program expert panel (7) recommends weight loss, exercise, and treatment of hypertension (particularly with angiotensin-converting enzyme inhibitors that improve insulin sensitivity) and dyslipidemia. These patients may experience a benefit from statin therapy even greater than that in patients with isolated increased concentrations of low-density lipoprotein cholesterol (LDL) (18). Mechanisms of Vascular Disease The mechanism for the disposition to atherosclerosis in patients with type 2 diabetes is becoming better defined. Traditional risk factors are more common in these patients than in healthy persons, but coronary artery disease events still exceed the expected rate by 50% (1). Table 2 shows the proposed pathophysiologic mechanisms that explain the independent risk factor status for coronary artery disease in diabetes. Table 2. Mechanisms of Increased Atherosclerosis in Diabetic Patients Endothelial Dysfunction Endothelial dysfunction is a precursor to and an effect of atherosclerosis. The vascular endothelium is a multifunctional organ system that resists thrombosis and atherogenesis and regulates blood flow by producing nitric oxide, the prime mediator of vascular reactivity. Diabetes impairs endothelial function (19) through several proposed mechanisms (Table 3). The clinical implications of endothelial dysfunction are not limited to increased atherosclerosis. Endothelial cells also help form collateral circulation, which is reduced in patients with diabetes (25) and may explain the increased infarct extension and congestive heart failure after myocardial infarction in these patients. Table 3. Mechanisms for Endothelial Dysfunction in Type 2 Diabetes Mellitus Dyslipidemia Dyslipidemia is a known risk factor for coronary atherosclerosis, and lipid abnormalities are common in diabetic patients. Elevated levels of small, dense LDL particles, low levels of high-density lipoprotein (HDL) cholesterol, and high triglyceride levels make up the condition known as diabetic dyslipidemia. In 1 study, patients with diabetic dyslipidemia had a worse prognosis than did those with isolated elevated LDL levels (26). However, dyslipidemia does not fully explain the increased incidence of coronary artery disease. At any lipid level, diabetic patients have more severe coronary disease than their nondiabetic counterparts, possibly because of an increased incidence of atherogenic lipid abnormalities, such as small, dense LDL particles. The presence of these particles (subclass phenotype B) is associated with a 3-fold increase in myocardial infarction (27) and is an important part of the metabolic syndrome (28). In addition, increased oxidation of LDL in diabetic patients has been associated with increased risk for coronary artery disease, possibly by promoting endothelial dysfunction. Oxidized LDL is cytotoxic to endothelium; it impairs endothelium-dependent vasodilation by inactivating nitric oxide (20) and causes endothelial disruption (29). Thrombogenesis The propensity for clotting is increased in patients with diabetes (30). The level of plasminogen activator inhibitor, which suppresses fibrinolysis, is elevated in the serum and atherectomy specimens of diabetic patients (31), and elevated levels are associated with an increased risk for myocardial infarction (32). In addition, increased concentrations of prothrombotic substances (such as tissue factor, fibrinogen, and factor VII) are found in diabetic patients (33). Clinically, this increased thrombogenicity improves with better glucose control (34). Platelet-dependent thrombosis has also been linked to an elevated glucose level (35), and diabetic patients have more platelet aggregation in response to platelet activation (36). This state of hyperaggregability of platelets may explain the enhanced efficacy of antiplatelet agents, such as glycoprotein IIb/IIIa inhibitors (37). Oxidative Stress On the molecular level, oxidative stress (a relative increase in oxygen free radicals) appears to play a role in diabetic atherogenesis (38). Hyperglycemia leads to increased production of reactive oxygen species and to nonenzymatic glycoxidation of proteins, which alters their structure and function. Ultimately, these altered proteins (known as advanced glycation end products) accumulate in patients with a chronically elevated glucose level (38). The effects of advanced glycation end products (through receptors for advanced glycation end products) cause increases in vascular permeability, procoagulant activity, adhesion molecule expression, and monocyte influx that contribute to vascular injury. Advanced glycation end products also affect dyslipidemia by altering LDL receptormediated clearance mechanisms (39). Autonomic Neuropathy Autonomic neuropathy, which leads to an increased propensity for malignant arrhythmia, is another possible mechanism for the high morbidity and mortality from coronary artery disease in diabetic patients. Sympathovagal imbalance from parasympathetic denervation occurs in 40% to 50% of patients (40). The resulting variation in areas of denervation in the myocardium may lead to arrhythmogenesis and sudden cardiac death (41). Preventive Management Modification of Risk Factors for Coronary Artery Disease and Glycemic Control Although it is intuitive that improved glucose control will decrease the risk for macrovascular events, this effect has not been conclusively demonstrated. The United Kingdom Prospective Diabetes Study 33 compared intensive glucose control with sulfonylureas or insulin and conventional care. A total of 3867 patients with newly diagnosed type 2 diabetes (median age, 54 years) were randomly allocated to receive intensive therapy with sulfonylureas or insulin and were followed for a mean of 10 years. Tight glycemic control reduced the incidence of microvascular but not macrovascular complications (42). Therefore, despite evidence of an increase in cardiovascular events with increasing blood glucose levels (43), no strong evidenc

Epicardial Fat: An Additional Measurement for Subclinical Atherosclerosis and Cardiovascular Risk Stratification?

OBJECTIVE To determine the ability of carotid intima-media thickness (CIMT) and coronary artery calcium score (CACS) to detect subclinical atherosclerosis in a young to middle-aged, low-risk, primary-prevention population. PATIENTS AND METHODS Patients aged 36 to 59 years who underwent determination of CIMT and CACS at our institution between May 1, 2004, and April 1, 2008, were included in the study. Those with diabetes mellitus or a history of coronary, peripheral, or cerebral vascular disease were excluded. Other information, such as Framingham risk score (FRS), was obtained by a review of clinical and laboratory data. RESULTS Of 118 patients, 89 (75%) had a CACS of zero and 94 (80%) were men; mean ± SD age was 48.9±5.7 years. The mean FRS of this group was 4.0; 86 patients (97%) were considered at low risk ( CONCLUSION Subclinical vascular disease can be detected by CIMT evaluation in young to middle-aged patients with a low FRS and a CACS of zero. These findings have important implications for vascular disease screening and the implementation of primary-prevention strategies.

Preoperative Dobutamine Stress Echocardiographic Findings and Subsequent Short-Term Adverse Cardiac Events After Orthotopic Liver Transplantation

BACKGROUND The value of epicardial adipose tissue (EAT) thickness as determined by echocardiography in cardiovascular risk assessment is not well understood. The aim of this study was to determine the associations between EAT thickness and Framingham risk score, carotid intima media thickness, carotid artery plaque, and computed tomographic coronary calcium score in a primary prevention population. METHODS Patients presenting for cardiovascular preventive care (n = 356) who underwent echocardiography as well as carotid artery ultrasound and/or coronary calcium scoring were included. RESULTS EAT thickness was weakly correlated with Framingham risk score. The prevalence of carotid plaque was significantly greater in those with EAT thickness ≥ 5.0 mm who either had low Framingham risk scores or had body mass indexes ≥ 25 kg/m(2), compared with those with EAT thickness <5.0 mm. No significant association between EAT thickness and carotid intima-media thickness or coronary calcium score existed. CONCLUSION EAT thickness ≥ 5.0 mm may identify an individual with a higher likelihood of having detectable carotid atherosclerosis.

Carotid ultrasound identifies high risk subclinical atherosclerosis in adults with low Framingham risk scores

Cardiovascular (CV) complications are the leading cause of non–graft‐related death in orthotopic liver transplant (OLT) patients. Pretransplant cardiac evaluation using dobutamine stress echocardiography (DSE) is commonly utilized for risk stratification of OLT candidates. To determine if clinical and echocardiographic variables identify patients with increased CV risk, we performed a retrospective chart review of all 284 patients that underwent OLT at our institution between June 1999 and August 2005. Of these patients, 157 had a DSE prior to their OLT. Serious adverse CV events occurring during surgery and up to 4 months post‐transplantation were defined as cardiac‐related death, myocardial infarction (MI), new heart failure, or asystole or unstable ventricular arrhythmia requiring acute treatment. Sixteen of 157 patients (10%) had an adverse CV event with 2 deaths. These included ventricular tachycardia (n = 8), asystole (n = 2), MI (n = 5), and new heart failure (n = 1). Nine of the 16 CV events occurred at the time of surgery (including both deaths), 5 occurred postoperatively, and 3 occurred after hospital discharge. Variables that correlated with increased CV events were inability during DSE to achieve >82% of the maximum predicted heart rate (22% versus 6%, P = 0.01), a peak rate pressure product during DSE of <16,333 (17% versus 5%, P = 0.02), and a Model for End‐Stage Liver Disease (MELD) score of >24 at the time of OLT. A multivariate model calculated from the DSE maximum achieved heart rate (MAHR) and MELD score (result = 3.78 + 0.07 MELD − 0.05 MAHR) identified a 47% risk for a value > 0 versus a 6% risk for a value < 0 (P < 0.001). In conclusion, the maximum heart rate achieved during DSE together with the MELD score may be a predictor of adverse CV events up to 4 months post‐OLT. A large prospective study is needed to more decisively support this conclusion. Liver Transpl 14:886–892, 2008.

A Summary and Critical Assessment of the 2013 ACC/AHA Guideline on the Treatment of Blood Cholesterol to Reduce Atherosclerotic Cardiovascular Disease Risk in Adults: Filling the Gaps

BACKGROUND Worldwide, cardiovascular (CV) disease remains the most common cause of morbidity and mortality. Although effective in predicting CV risk in select populations, the Framingham risk score (FRS) fails to identify many young individuals who experience premature CV events. Accordingly, the aim of this study was to determine the prevalence of high-risk carotid intima-media thickness (CIMT) or plaque, a marker of atherosclerosis and predictor of CV events, in young asymptomatic individuals with low and intermediate FRS (<2% annualized event rate) using the carotid ultrasound protocol recommended by the American Society of Echocardiography and the Society of Vascular Medicine. METHODS Individuals aged < or = 65 years not taking statins and without diabetes mellitus or histories of coronary artery disease underwent CIMT and plaque examination for primary prevention. Clinical variables including lipid values, family history of premature coronary artery disease, and FRS and subsequent pharmacotherapy recommendations were retrospectively collected for statistical analysis. RESULTS Of 441 subjects (mean age, 49.7 + or - 7.9 years), 184 (42%; 95% confidence interval, 37.3%-46.5%) had high-risk carotid ultrasound findings (CIMT > or = 75th percentile adjusted for age, gender, and race or presence of plaque). Of those with the lowest FRS of < or =5% (n = 336) (mean age, 48.0 + or - 7.6 years; mean FRS, 2.5 + or - 1.5%), 127 (38%; 95% confidence interval, 32.6%-43.0%) had high-risk carotid ultrasound findings. For individuals with FRS < or = 5% and high-risk carotid ultrasound findings (n = 127; mean age, 47.3 + or - 8.1 years; mean FRS, 2.5 + or - 1.5%), lipid-lowering therapy was recommended by their treating physicians in 77 (61%). CONCLUSIONS Thirty-eight percent of asymptomatic young to middle-aged individuals with FRS < or = 5% have abnormal carotid ultrasound findings associated with increased risk for CV events. Pharmacologic therapy for CV prevention was recommended in the majority of these individuals. The lack of radiation exposure, relatively low cost, and ability to detect early-stage atherosclerosis suggest that carotid ultrasound for CIMT and plaque detection should continue to be explored as a primary tool for CV risk stratification in young to middle-aged adults with low FRS.

Transient left ventricular apical ballooning syndrome (Takotsubo cardiomyopathy) following orthotopic liver transplantation

The American College of Cardiology/American Heart Association (ACC/AHA) Task Force on Practice Guidelines has recently released the new cholesterol treatment guideline. This update was based on a systematic review of the evidence and replaces the previous guidelines from 2002 that were widely accepted and implemented in clinical practice. The new cholesterol treatment guideline emphasizes matching the intensity of statin treatment to the level of atherosclerotic cardiovascular disease (ASCVD) risk and replaces the old paradigm of pursuing low-density lipoprotein cholesterol targets. The new guideline also emphasizes the primacy of the evidence base for statin therapy for ASCVD risk reduction and lists several patient groups that will not benefit from statin treatment despite their high cardiovascular risk, such as those with heart failure (New York Heart Association class II-IV) and patients undergoing hemodialysis. The guideline has been received with mixed reviews and significant controversy. Because of the evidence-based nature of the guideline, there is room for several questions and uncertainties on when and how to use lipid-lowering therapy in clinical practice. The goal of the Mayo Clinic Task Force in the assessment, interpretation, and expansion of the ACC/AHA cholesterol treatment guideline is to address gaps in information and some of the controversial aspects of the newly released cholesterol management guideline using additional sources of evidence and expert opinion as needed to guide clinicians on key aspects of ASCVD risk reduction.