Are you sure it's an artifact? Artifact detection and uncertainty quantification in histological images

Modern cancer diagnostics involves extracting tissue specimens from suspicious areas and conducting histotechnical procedures to prepare a digitized glass slide, called Whole Slide Image (WSI), for further examination. These procedures frequently introduce different types of artifacts in the obtained WSI, and histological artifacts might influence Computational Pathology (CPATH) systems further down to a diagnostic pipeline if not excluded or handled. Deep Convolutional Neural Networks (DCNNs) have achieved promising results for the detection of some WSI artifacts, however, they do not incorporate uncertainty in their predictions. This paper proposes an uncertainty-aware Deep Kernel Learning (DKL) model to detect blurry areas and folded tissues, two types of artifacts that can appear in WSIs. The proposed probabilistic model combines a CNN feature extractor and a sparse Gaussian Processes (GPs) classifier, which improves the performance of current state-of-the-art artifact detection DCNNs and provides uncertainty estimates. We achieved 0.996 and 0.938 F1 scores for blur and folded tissue detection on unseen data, respectively. In extensive experiments, we validated the DKL model on unseen data from external independent cohorts with different staining and tissue types, where it outperformed DCNNs. Interestingly, the DKL model is more confident in the correct predictions and less in the wrong ones. The proposed DKL model can be integrated into the preprocessing pipeline of CPATH systems to provide reliable predictions and possibly serve as a quality control tool.

Artificial Intelligence in Digital Pathology for Bladder Cancer: Hype or Hope? A Systematic Review

Bladder cancer (BC) diagnosis and prediction of prognosis are hindered by subjective pathological evaluation, which may cause misdiagnosis and under-/over-treatment. Computational pathology (CPATH) can identify clinical outcome predictors, offering an objective approach to improve prognosis. However, a systematic review of CPATH in BC literature is lacking. Therefore, we present a comprehensive overview of studies that used CPATH in BC, analyzing 33 out of 2285 identified studies. Most studies analyzed regions of interest to distinguish normal versus tumor tissue and identify tumor grade/stage and tissue types (e.g., urothelium, stroma, and muscle). The cell's nuclear area, shape irregularity, and roundness were the most promising markers to predict recurrence and survival based on selected regions of interest, with >80% accuracy. CPATH identified molecular subtypes by detecting features, e.g., papillary structures, hyperchromatic, and pleomorphic nuclei. Combining clinicopathological and image-derived features improved recurrence and survival prediction. However, due to the lack of outcome interpretability and independent test datasets, robustness and clinical applicability could not be ensured. The current literature demonstrates that CPATH holds the potential to improve BC diagnosis and prediction of prognosis. However, more robust, interpretable, accurate models and larger datasets-representative of clinical scenarios-are needed to address artificial intelligence's reliability, robustness, and black box challenge.

Machine learning algorithms vs. thresholding to segment ischemic regions in patients with acute ischemic stroke

OBJECTIVE

Computed tomography (CT) scan is a fast and widely used modality for early assessment in patients with symptoms of a cerebral ischemic stroke. CT perfusion (CTP) is often added to the protocol and is used by radiologists for assessing the severity of the stroke. Standard parametric maps are calculated from the CTP datasets. Based on parametric value combinations, ischemic regions are separated into presumed infarct core (irreversibly damaged tissue) and penumbra (tissue-at-risk). Different thresholding approaches have been suggested to segment the parametric maps into these areas. The purpose of this study is to compare fully-automated methods based on machine learning and thresholding approaches to segment the hypoperfused regions in patients with ischemic stroke.

METHODS

We test two different architectures with three mainstream machine learning algorithms. We use parametric maps, as input features, and manual annotations made by two expert neuroradiologists as ground truth.

RESULTS

The best results are produced with random forest (RF) and Single-Step approach; we achieve an average Dice coefficient of 0.68 and 0.26, respectively for penumbra and core, for the three groups analysed. We also achieve an average in volume difference of 25.1ml for penumbra and 7.8ml for core.

CONCLUSIONS

Our best RF-based method outperforms the classical thresholding approaches, to segment both the ischemic regions in a group of patients regardless of the severity of vessel occlusion.

SIGNIFICANCE

A correct visualization of the ischemic regions will guide treatment decision better.

Fetal heart rate development during labour

BACKGROUND

Fresh stillbirths (FSB) and very early neonatal deaths (VEND) are important global challenges with 2.6 million deaths annually. The vast majority of these deaths occur in low- and low-middle income countries. Assessment of the fetal well-being during pregnancy, labour, and birth is normally conducted by monitoring the fetal heart rate (FHR). The heart rate of newborns is reported to increase shortly after birth, but a corresponding trend in how FHR changes just before birth for normal and adverse outcomes has not been studied. In this work, we utilise FHR measurements collected from 3711 labours from a low and low-middle income country to study how the FHR changes towards the end of the labour. The FHR development is also studied in groups defined by the neonatal well-being 24 h after birth.

METHODS

A signal pre-processing method was applied to identify and remove time periods in the FHR signal where the signal is less trustworthy. We suggest an analysis framework to study the FHR development using the median FHR of all measured heart rates within a 10-min window. The FHR trend is found for labours with a normal outcome, neonates still admitted for observation and perinatal mortality, i.e. FSB and VEND. Finally, we study how the spread of the FHR changes over time during labour.

RESULTS

When studying all labours, there is a drop in median FHR from 134 beats per minute (bpm) to 119 bpm the last 150 min before birth. The change in FHR was significant ([Formula: see text]) using Wilcoxon signed-rank test. A drop in median FHR as well as an increased spread in FHR is observed for all defined outcome groups in the same interval.

CONCLUSION

A significant drop in FHR the last 150 min before birth is seen for all neonates with a normal outcome or still admitted to the NCU at 24 h after birth. The observed earlier and larger drop in the perinatal mortality group may indicate that they struggle to endure the physical strain of labour, and that an earlier intervention could potentially save lives. Due to the low amount of data in the perinatal mortality group, a larger dataset is required to validate the drop for this group.

A Multiscale Approach for Whole-Slide Image Segmentation of five Tissue Classes in Urothelial Carcinoma Slides

In pathology labs worldwide, we see an increasing number of tissue samples that need to be assessed without the same increase in the number of pathologists. Computational pathology, where digital scans of histological samples called whole-slide images (WSI) are processed by computational tools, can be of help for the pathologists and is gaining research interests. Most research effort has been given to classify slides as being cancerous or not, localization of cancerous regions, and to the “big-four” in cancer: breast, lung, prostate, and bowel. Urothelial carcinoma, the most common form of bladder cancer, is expensive to follow up due to a high risk of recurrence, and grading systems have a high degree of inter- and intra-observer variability. The tissue samples of urothelial carcinoma contain a mixture of damaged tissue, blood, stroma, muscle, and urothelium, where it is mainly muscle and urothelium that is diagnostically relevant. A coarse segmentation of these tissue types would be useful to i) guide pathologists to the diagnostic relevant areas of the WSI, and ii) use as input in a computer-aided diagnostic (CAD) system. However, little work has been done on segmenting tissue types in WSIs, and on computational pathology for urothelial carcinoma in particular. In this work, we are using convolutional neural networks (CNN) for multiscale tile-wise classification and coarse segmentation, including both context and detail, by using three magnification levels: 25x, 100x, and 400x. 28 models were trained on weakly labeled data from 32 WSIs, where the best model got an F1-score of 96.5% across six classes. The multiscale models were consistently better than the single-scale models, demonstrating the benefit of combining multiple scales. No tissue-class ground-truth for complete WSIs exist, but the best models were used to segment seven unseen WSIs where the results were manually inspected by a pathologist and are considered as very promising.

Automatic identification of stimulation activities during newborn resuscitation using ECG and accelerometer signals

BACKGROUND AND OBJECTIVE

Early neonatal death is a worldwide challenge with 1 million newborn deaths every year. The primary cause of these deaths are complications during labour and birth asphyxia. The majority of these newborns could have been saved with adequate resuscitation at birth. Newborn resuscitation guidelines recommend immediate drying, stimulation, suctioning if indicated, and ventilation of non-breathing newborns. A system that will automatically detect and extract time periods where different resuscitation activities are performed, would be highly beneficial to evaluate what resuscitation activities that are improving the state of the newborn, and if current guidelines are good and if they are followed. The potential effects of especially stimulation are not very well documented as it has been difficult to investigate through observations. In this paper the main objective is to identify stimulation activities, regardless if the state of the newborn is changed or not, and produce timelines of the resuscitation episode with the identified stimulations.

METHODS

Data is collected by utilizing a new heart rate device, NeoBeat, with dry-electrode ECG and accelerometer sensors placed on the abdomen of the newborn. We propose a method, NBstim, based on time domain and frequency domain features from the accelerometer signals and ECG signals from NeoBeat, to detect time periods of stimulation. NBstim use causal features from a gliding window of the signals, thus it can potentially be used in future realtime systems. A high performing feature subset is found using feature selection. System performance is computed using a leave-one-out cross-validation and compared with manual annotations.

RESULTS

The system achieves an overall accuracy of 90.3% when identifying regions with stimulation activities.

CONCLUSION

The performance indicates that the proposed NBstim, used with signals from the NeoBeat can be used to determine when stimulation is performed. The provided activity timelines, in combination with the status of the newborn, for example the heart rate, at different time points, can be studied further to investigate both the time spent and the effect of different newborn resuscitation parameters.

Activity Recognition From Newborn Resuscitation Videos

OBJECTIVE

Birth asphyxia is one of the leading causes of neonatal deaths. A key for survival is performing immediate and continuous quality newborn resuscitation. A dataset of recorded signals during newborn resuscitation, including videos, has been collected in Haydom, Tanzania, and the aim is to analyze the treatment and its effect on the newborn outcome. An important step is to generate timelines of relevant resuscitation activities, including ventilation, stimulation, suction, etc., during the resuscitation episodes.

METHODS

We propose a two-step deep neural network system, ORAA-net, utilizing low-quality video recordings of resuscitation episodes to do activity recognition during newborn resuscitation. The first step is to detect and track relevant objects using Convolutional Neural Networks (CNN) and post-processing, and the second step is to analyze the proposed activity regions from step 1 to do activity recognition using 3D CNNs.

RESULTS

The system recognized the activities newborn uncovered, stimulation, ventilation and suction with a mean precision of 77.67%, a mean recall of 77,64%, and a mean accuracy of 92.40%. Moreover, the accuracy of the estimated number of Health Care Providers (HCPs) present during the resuscitation episodes was 68.32%.

CONCLUSION

The results indicate that the proposed CNN-based two-step ORAA-net could be used for object detection and activity recognition in noisy low-quality newborn resuscitation videos.

SIGNIFICANCE

A thorough analysis of the effect the different resuscitation activities have on the newborn outcome could potentially allow us to optimize treatment guidelines, training, debriefing, and local quality improvement in newborn resuscitation.

State transition modeling of complex monitored health data

This article considers the analysis of complex monitored health data, where often one or several signals are reflecting the current health status that can be represented by a finite number of states, in addition to a set of covariates. In particular, we consider a novel application of a non-parametric state intensity regression method in order to study time-dependent effects of covariates on the state transition intensities. The method can handle baseline, time varying as well as dynamic covariates. Because of the non-parametric nature, the method can handle different data types and challenges under minimal assumptions. If the signal that is reflecting the current health status is of continuous nature, we propose the application of a weighted median and a hysteresis filter as data pre-processing steps in order to facilitate robust analysis. In intensity regression, covariates can be aggregated by a suitable functional form over a time history window. We propose to study the estimated cumulative regression parameters for different choices of the time history window in order to investigate short- and long-term effects of the given covariates. The proposed framework is discussed and applied to resuscitation data of newborns collected in Tanzania.

Object Detection During Newborn Resuscitation Activities

OBJECTIVE

Birth asphyxia is a major newborn mortality problem in low-resource countries. International guideline provides treatment recommendations; however, the importance and effect of the different treatments are not fully explored. The available data are collected in Tanzania, during newborn resuscitation, for analysis of the resuscitation activities and the response of the newborn. An important step in the analysis is to create activity timelines of the episodes, where activities include ventilation, suction, stimulation, etc. Methods: The available recordings are noisy real-world videos with large variations. We propose a two-step process in order to detect activities possibly overlapping in time. The first step is to detect and track the relevant objects, such as bag-mask resuscitator, heart rate sensors, etc., and the second step is to use this information to recognize the resuscitation activities. The topic of this paper is the first step, and the object detection and tracking are based on convolutional neural networks followed by post processing.

RESULTS

The performance of the object detection during activities were 96.97% (ventilations), 100% (attaching/removing heart rate sensor), and 75% (suction) on a test set of 20 videos. The system also estimate the number of health care providers present with a performance of 71.16%.

CONCLUSION

The proposed object detection and tracking system provides promising results in noisy newborn resuscitation videos.

SIGNIFICANCE

This is the first step in a thorough analysis of newborn resuscitation episodes, which could provide important insight about the importance and effect of different newborn resuscitation activities.

Automatic Estimation of Coronary Blood Flow Velocity Step 1 for Developing a Tool to Diagnose Patients With Micro-Vascular Angina Pectoris

Aim: Our aim was to automatically estimate the blood velocity in coronary arteries using cine X-ray angiographic sequence. Estimating the coronary blood velocity is a key approach in investigating patients with angina pectoris and no significant coronary artery disease. Blood velocity estimation is central in assessing coronary flow reserve. Methods and Results: A multi-step automatic method for blood flow velocity estimation based on the information extracted solely from the cine X-ray coronary angiography sequence obtained by invasive selective coronary catheterization was developed. The method includes (1) an iterative process of segmenting coronary arteries modeling and removing the heart motion using a non-rigid registration, (2) measuring the area of the segmented arteries in each frame, (3) fitting the measured sequence of areas with a 7° polynomial to find start and stop time of dye propagation, and (4) estimating the blood flow velocity based on the time of the dye propagation and the length of the artery-tree. To evaluate the method, coronary angiography recordings from 21 patients with no obstructive coronary artery disease were used. In addition, coronary flow velocity was measured in the same patients using a modified transthoracic Doppler assessment of the left anterior descending artery. We found a moderate but statistically significant correlation between flow velocity assessed by trans thoracic Doppler and the proposed method applying both Spearman and Pearson tests. Conclusion: Measures of coronary flow velocity using a novel fully automatic method that utilizes the information from the X-ray coronary angiographic sequence were statistically significantly correlated to measurements obtained with transthoracic Doppler recordings.

ECG-Based Classification of Resuscitation Cardiac Rhythms for Retrospective Data Analysis

OBJECTIVE

There is a need to monitor the heart rhythm in resuscitation to improve treatment quality. Resuscitation rhythms are categorized into: ventricular tachycardia (VT), ventricular fibrillation (VF), pulseless electrical activity (PEA), asystole (AS), and pulse-generating rhythm (PR). Manual annotation of rhythms is time-consuming and infeasible for large datasets. Our objective was to develop ECG-based algorithms for the retrospective and automatic classification of resuscitation cardiac rhythms.

METHODS

The dataset consisted of 1631 3-s ECG segments with clinical rhythm annotations, obtained from 298 out-of-hospital cardiac arrest patients. In total, 47 wavelet- and time-domain-based features were computed from the ECG. Features were selected using a wrapper-based feature selection architecture. Classifiers based on Bayesian decision theory, k-nearest neighbor, k-local hyperplane distance nearest neighbor, artificial neural network (ANN), and ensemble of decision trees were studied.

RESULTS

The best results were obtained for ANN classifier with Bayesian regularization backpropagation training algorithm with 14 features, which forms the proposed algorithm. The overall accuracy for the proposed algorithm was 78.5%. The sensitivities (and positive-predictive-values) for AS, PEA, PR, VF, and VT were 88.7% (91.0%), 68.9% (70.4%), 65.9% (69.0%), 86.2% (83.8%), and 78.8% (72.9%), respectively.

CONCLUSIONS

The results demonstrate that it is possible to classify resuscitation cardiac rhythms automatically, but the accuracy for the organized rhythms (PEA and PR) is low.

SIGNIFICANCE

We have made an important step toward making classification of resuscitation rhythms more efficient in the sense of minimal feedback from human experts.

Chest compression rate measurement from smartphone video

Background

Out-of-hospital cardiac arrest is a life threatening situation where the first person performing cardiopulmonary resuscitation (CPR) most often is a bystander without medical training. Some existing smartphone apps can call the emergency number and provide for example global positioning system (GPS) location like Hjelp 113-GPS App by the Norwegian air ambulance. We propose to extend functionality of such apps by using the built in camera in a smartphone to capture video of the CPR performed, primarily to estimate the duration and rate of the chest compression executed, if any.

Methods

All calculations are done in real time, and both the caller and the dispatcher will receive the compression rate feedback when detected. The proposed algorithm is based on finding a dynamic region of interest in the video frames, and thereafter evaluating the power spectral density by computing the fast fourier transform over sliding windows. The power of the dominating frequencies is compared to the power of the frequency area of interest. The system is tested on different persons, male and female, in different scenarios addressing target compression rates, background disturbances, compression with mouth-to-mouth ventilation, various background illuminations and phone placements. All tests were done on a recording Laerdal manikin, providing true compression rates for comparison.

Results

Overall, the algorithm is seen to be promising, and it manages a number of disturbances and light situations. For target rates at 110 cpm, as recommended during CPR, the mean error in compression rate (Standard dev. over tests in parentheses) is 3.6 (0.8) for short hair bystanders, and 8.7 (6.0) including medium and long haired bystanders.

Conclusions

The presented method shows that it is feasible to detect the compression rate of chest compressions performed by a bystander by placing the smartphone close to the patient, and using the built-in camera combined with a video processing algorithm performed real-time on the device.

Automatic Detection and Parameterization of Manual Bag-Mask Ventilation on Newborns

OBJECTIVES

Birth asphyxia is a condition where a fetus suffers from lack of oxygen during birth. Intervention by manual ventilation should start within one minute after birth. Bag-mask resuscitators are commonly used in situations where ventilation is provided by a single health care worker. Due to a high complexity of interactions between physiological conditions of the newborns and the clinical treatment, the recommendations for bag-mask ventilation of infants remains controversial. The purpose of this paper is to illustrate the processing and parameterization of ventilation signals recorded from the Laerdal newborn resuscitation monitor into meaningful data.

METHODS

Basic signal processing approaches are applied on various signal channels (airway pressure, flow, CO ₂, and ECG) to detect events related to ventilation activities.

RESULTS

Different types of events are detected and parameterized to describe the characteristics of ventilation procedure.

CONCLUSIONS

Efficient detection algorithms as well as parameterization of ventilation events could be useful for retrospective analysis of resuscitation data, for example, by finding the association between different ventilation parameters and positive responses of newborns.

SIGNIFICANCE

Information about ventilation events and ventilation parameters could potentially be useful during a resuscitation situation by giving immediate feedback to the health care provider.

The relationship between transmurality of ischemic scars and the heart rate of ventricular tachycardia

AIMS

The relationship between the heart rate of ventricular tachycardia (VT) and the transmurality of ischemic scars was assessed by a new semiautomatic coordinate-based analysis of late gadolinium-enhanced cardiac magnetic resonance (LGE-CMR) images.

METHODS AND RESULTS

Twenty patients assessed by LGE-CMR before implantation of implantable cardioverter defibrillator (ICD) with verified VT during the first year following ICD implantation were included. Scar was defined by pixels with a signal intensity ≥ 50% of maximum signal intensity. All pixels were assigned a coordinate position between endo- and epicardium (λ) and the angle of the heart axis (φ). Based upon the λ and φ values, multiple scar features were computed for all scarred areas. These features were correlated to VT heart rate across the complete range of transmurality. The strongest correlation with univariate regression was found between VT heart rate and the sum of transmurality when the maximum transmurality of these features was ≥ 90% (R-square = 0.47). In multiple regressions analysis, the strongest relationship with VT heart rate was found with a maximum transmurality ≥ 90% and by a combination of scar size, transmurality, and endocardial extent of infarction (R-square = 0.64).

CONCLUSION

Transmurality is the strongest predictor of VT heart rate both in univariate and multivariate models. The strongest relationships were found at a transmurality level > 90%.

Cardiac magnetic resonance image-based classification of the risk of arrhythmias in post-myocardial infarction patients

INTRODUCTION

Patients surviving myocardial infarction (MI) can be divided into high and low arrhythmic risk groups. Distinguishing between these two groups is of crucial importance since the high-risk group has been shown to benefit from implantable cardioverter defibrillator insertion; a costly surgical procedure with potential complications and no proven advantages for the low-risk group. Currently, markers such as left ventricular ejection fraction and myocardial scar size are used to evaluate arrhythmic risk.

METHODS

In this paper, we propose quantitative discriminative features extracted from late gadolinium enhanced cardiac magnetic resonance images of post-MI patients, to distinguish between 20 high-risk and 34 low-risk patients. These features include size, location, and textural information concerning the scarred myocardium. To evaluate the discriminative power of the proposed features, we used several built-in classification schemes from matrix laboratory (MATLAB) and Waikato environment for knowledge analysis (WEKA) software, including k-nearest neighbor (k-NN), support vector machine (SVM), decision tree, and random forest.

RESULTS

In Experiment 1, the leave-one-out cross-validation scheme is implemented in MATLAB to classify high- and low-risk groups with a classification accuracy of 94.44%, and an AUC of 0.965 for a feature combination that captures size, location and heterogeneity of the scar. In Experiment 2 with the help of WEKA, nested cross-validation is performed with k-NN, SVM, adjusting decision tree and random forest classifiers to differentiate high-risk and low-risk patients. SVM classifier provided average accuracy of 92.6%, and AUC of 0.921 for a feature combination capturing location and heterogeneity of the scar. Experiment 1 and Experiment 2 show that textural features from the scar are important for classification and that localization features provide an additional benefit.

CONCLUSION

These promising results suggest that the discriminative features introduced in this paper can be used by medical professionals, or in automatic decision support systems, along with the recognized risk markers, to improve arrhythmic risk stratification in post-MI patients.

Probability mapping of scarred myocardium using texture and intensity features in CMR images

Background

The myocardium exhibits heterogeneous nature due to scarring after Myocardial Infarction (MI). In Cardiac Magnetic Resonance (CMR) imaging, Late Gadolinium (LG) contrast agent enhances the intensity of scarred area in the myocardium.

Methods

In this paper, we propose a probability mapping technique using Texture and Intensity features to describe heterogeneous nature of the scarred myocardium in Cardiac Magnetic Resonance (CMR) images after Myocardial Infarction (MI). Scarred tissue and non-scarred tissue are represented with high and low probabilities, respectively. Intermediate values possibly indicate areas where the scarred and healthy tissues are interwoven. The probability map of scarred myocardium is calculated by using a probability function based on Bayes rule. Any set of features can be used in the probability function.

Results

In the present study, we demonstrate the use of two different types of features. One is based on the mean intensity of pixel and the other on underlying texture information of the scarred and non-scarred myocardium. Examples of probability maps computed using the mean intensity of pixel and the underlying texture information are presented. We hypothesize that the probability mapping of myocardium offers alternate visualization, possibly showing the details with physiological significance difficult to detect visually in the original CMR image.

Conclusion

The probability mapping obtained from the two features provides a way to define different cardiac segments which offer a way to identify areas in the myocardium of diagnostic importance (like core and border areas in scarred myocardium).

Exploratory data analysis of image texture and statistical features on myocardium and infarction areas in cardiac magnetic resonance images

The cardiac magnetic resonance (CMR) images from a group of patients with myocardial scars and implanted cardioverter-defibrillator (ICD) are divided into a group with low risk of arrhythmias (late incidents) and a group with high risk of arrhythmias (early incidents). Several hundred quantitative features describing sizes, statistics and textures of the segmented and defined areas of the images are computed from manually segmented images in an exploratory analysis. The method used to determine decision regions to discriminate the patients with low risk of arrhythmias from the patient with high risk of arrhythmias is a maximum likelihood estimation based Bayes classifiers described in [1]. The results presented can be interpreted as hypothesis of which features, and combinations of features, that might have discriminative power. A major hypothesis that arises is that there are important textural information in the scarred and non-scarred areas.

Segmentation of scarred and non-scarred myocardium in LG enhanced CMR images using intensity-based textural analysis

The Late Gadolinium (LG) enhancement in Cardiac Magnetic Resonance (CMR) imaging is used to increase the intensity of scarred area in myocardium for thorough examination. Automatic segmentation of scar is important because scar size is largely responsible in changing the size, shape and functioning of left ventricle and it is a preliminary step required in exploring the information present in scar. We have proposed a new technique to segment scar (infarct region) from non-scarred myocardium using intensity-based texture analysis. Our new technique uses dictionary-based texture features and dc-values to segment scarred and non-scarred myocardium using Maximum Likelihood Estimator (MLE) based Bayes classification. Texture analysis aided with intensity values gives better segmentation of scar from myocardium with high sensitivity and specificity values in comparison to manual segmentation by expert cardiologists.

Watershed segmentation of detected masses in digital mammograms

A method for segmentation of detected masses in digital mammograms is introduced. The method is based on gray scale mathematical morphology. In a preprocessing step, image enhancement based on a local histogram technique is applied, followed by a morphological smoothing operation. The watershed transform is then applied to the gradient of the smoothed image resulting in segmented regions. A good segmentation is important in order to be able to extract useful feature measures from the segmented regions. These feature measures can be input to a classifier which classifies each region as either a mass or a false detection. Initial experiments have been performed using mammograms from the MIAS database. Results of the experimental study indicate that our scheme can provide useful contour extraction for mass structures.

Dictionary learning algorithms for sparse representation

Algorithms for data-driven learning of domain-specific overcomplete dictionaries are developed to obtain maximum likelihood and maximum a posteriori dictionary estimates based on the use of Bayesian models with concave/Schur-concave (CSC) negative log priors. Such priors are appropriate for obtaining sparse representations of environmental signals within an appropriately chosen (environmentally matched) dictionary. The elements of the dictionary can be interpreted as concepts, features, or words capable of succinct expression of events encountered in the environment (the source of the measured signals). This is a generalization of vector quantization in that one is interested in a description involving a few dictionary entries (the proverbial "25 words or less"), but not necessarily as succinct as one entry. To learn an environmentally adapted dictionary capable of concise expression of signals generated by the environment, we develop algorithms that iterate between a representative set of sparse representations found by variants of FOCUSS and an update of the dictionary using these sparse representations. Experiments were performed using synthetic data and natural images. For complete dictionaries, we demonstrate that our algorithms have improved performance over other independent component analysis (ICA) methods, measured in terms of signal-to-noise ratios of separated sources. In the overcomplete case, we show that the true underlying dictionary and sparse sources can be accurately recovered. In tests with natural images, learned overcomplete dictionaries are shown to have higher coding efficiency than complete dictionaries; that is, images encoded with an overcomplete dictionary have both higher compression (fewer bits per pixel) and higher accuracy (lower mean square error).