Michael Berks

Detecting and classifying linear structures in mammograms using random forests

Detecting and classifying curvilinear structure is important in many image interpretation tasks. We focus on the challenging problem of detecting such structure in mammograms and deciding whether it is normal or abnormal. We adopt a discriminative learning approach based on a Dual-Tree Complex Wavelet representation and random forest classification. We present results of a quantitative comparison of our approach with three leading methods from the literature and with learning-based variants of those methods. We show that our new approach gives significantly better results than any of the other methods, achieving an area under the ROC curve A(z) = 0.923 for curvilinear structure detection, and A(z) = 0.761 for distinguishing between normal and abnormal structure (spicules). A detailed analysis suggests that some of the improvement is due to discriminative learning, and some due to the DT-CWT representation, which provides local phase information and good angular resolution.

An Automated System for Detecting and Measuring Nailfold Capillaries

Nailfold capillaroscopy is an established qualitative technique in the assessment of patients displaying Raynauds phenomenon. We describe a fully automated system for extracting quantitative biomarkers from capillaroscopy images, using a layered machine learning approach. On an unseen set of 455 images, the system detects and locates individual capillaries as well as human experts, and makes measurements of vessel morphology that reveal statistically significant differences between patients with (relatively benign) primary Raynauds phenomenon, and those with potentially life-threatening systemic sclerosis.

Statistical Appearance Models of Mammographic Masses

We present a method for building generative statistical appearance models of mammographic masses. We address several key issues that limited the performance of previous methods. In particular, we use MDL optimization to generate more compact shape correspondences; we describe a technique for the accurate estimation of the background tissue on which a mass is superimposed; and we highlight the importance of choosing suitable weighting between shape, texture and scale components in the final combined model. Improvements in the ability of the model to characterize a set of 101 mammographic masses are quantified using leave-one-out testing, showing a reduction in mean square error per pixel from 3.109 using a previous method to 1.262 using the new appearance model.

Volumetric and area-based breast density measurement in the predicting risk of cancer at screening (PROCAS) study

Mammographic density, defined as the proportion of the breast area in a mammogram that contains fibroglandular tissue, is associated with risk of breast cancer. However, measures of mammographic density are subject to variation in the underlying imaging process and in the assessments of observers. Automatic volumetric measures of breast density remove much of this variability, but their association with risk is less well established. We present density measurements produced using area-based visual analogue scales (VAS) and by volumetric assessment software (QuantraTM, Hologic Inc.) in the PROCAS study. The distributions of VAS scores (n = 22 327) and volumetric quantities (n = 11 653) are given, as are their relationships for subjects with results by both (n = 11 096), but these are not directly comparable as one is area-based and the other volumetric. Inter-observer variability in visual area-based estimation is examined by a scatter plot matrix.

Visual assessment of density in digital mammograms

This study compares visual assessment of density on full field digital mammograms using visual analogue scales (VAS) and written percentages Fifty normal digital screening mammograms were selected at random Nine readers viewed the images on two occasions, firstly indicating density on a VAS and then estimating the percentage of dense tissue in the breast Although the two methods were correlated, the degree of agreement between the density estimates varied considerably from reader to reader More experienced readers used a wider range of values, and inter-observer variability for both methods was higher for these readers The greatest difference between the methods was in mammograms with a mixed fatty-glandular appearance (density between 55% and 75%) Both methods are quick and convenient, although these results demonstrate a need for training to ensure they are used consistently by readers of different degrees of experience.

Feasibility and acceptability of stepwedge-based density measurement

A link between increased breast density, as visualised in mammograms, and increased risk of developing breast cancer has been established. Recently, a number of objective, quantitative methods for measuring breast density have been described. One such method requires a calibration object to be imaged alongside the breast. However, it is important that this should not interfere with the routine imaging process. In this paper, we investigate the amount of space in mammographic images which is not currently occupied by the breast or existing patient labels and markers, and which would therefore be available for imaging an additional calibration device. We do this with a view to estimating the likelihood of failure of the method, and also to determining whether, without detriment to the imaging process, a device could be permanently fixed to the breast support platform. We also examine the impact of markers attached to the compression plate on the visibility of breast tissue. The results show that our existing calibration device may be used successfully without interfering with the routine imaging process, although permanently fixing such a device may present problems in a small minority of cases, and we demonstrate that the number of cases which would fail can be reduced by using a smaller stepwedge.

Simulating nailfold capillaroscopy sequences to evaluate algorithms for blood flow estimation

The effects of systemic sclerosis (SSc) - a disease of the connective tissue causing blood flow problems that can require amputation of the fingers - can be observed indirectly by imaging the capillaries at the nailfold, though taking quantitative measures such as blood flow to diagnose the disease and monitor its progression is not easy. Optical flow algorithms may be applied, though without ground truth (i.e. known blood flow) it is hard to evaluate their accuracy. We propose an image model that generates realistic capillaroscopy videos with known flow, and use this model to quantify the effect of flow rate, cell density and contrast (among others) on estimated flow. This resource will help researchers to design systems that are robust under real-world conditions.

Modelling structural deformations in mammographic tissue using the dual-tree complex wavelet

The appearance of breast tissue in mammograms is altered by the presence of a malignant mass Existing synthesis methods have not addressed this structural deformation We aim to use a set of mass background images that display altered breast tissue to simulate such deformations in regions of digital mammograms previously showing no signs of disease Regions are decomposed using the dual-tree complex wavelet transform (DT-CWT) to obtain a richer representation of local structure than provided by image grey-levels alone Synthesis is achieved by modifying the high-frequency DT-CWT coefficients of normal regions to match those in mass backgrounds Three methods for completing this task are described The results, advantages and current limitations of the methods are discussed.

Improved Diagnosis of Systemic Sclerosis Using Nailfold Capillary Flow

Nailfold capillaroscopy (NC) allows non-invasive imaging of systemic sclerosis (SSc) related microvascular disease. We have developed a state-of-the-art NC system that enables fast, panoramic imaging of the whole nailfold at high-magnification, and incorporates novel software to make fully automated estimates of capillary structure and blood flow velocity. We present the first results of a study in which 50 patients with SSc, 12 with primary Raynauds phenomenon (PRP) and 50 healthy controls (HC) were imaged using the new system, and show that a combined model of capillary measurements strongly separates SSc from HC/PRP (ROC \(A_z\)=0.93). Including capillary flow improves model performance, suggesting flow provides complementary information to capillary structure for diagnosing SSc.

Synthesising malignant breast masses in normal mammograms

Using mammograms in which signs of breast cancer have been synthesised overcomes the problem of obtaining a sufficiently large volume of real data with known ground truth for training and test purposes This paper describes a fully automated method for generating synthetic spiculated masses Statistical methods are used to model the appearance and location of a training set of real masses and their effect on surrounding breast tissue The models are then used to synthesise the appearance of a malignant mass in an otherwise normal mammogram By virtue of using generative statistical models, the synthesis process can be fully automated In an observer study in which 10 expert mammogram readers attempted to distinguish between synthetic masses generated by the method and real masses, we report an area Az = 0.70±0.09 under the receiver operating characteristic.