Helene Schulerud

A review of caveats in statistical nuclear image analysis

A large body of the published literature in nuclear image analysis do not evaluate their findings on an independent data set. Hence, if several features are evaluated on a limited data set over‐optimistic results are easily achieved. In order to find features that separate different outcome classes of interest, statistical evaluation of the nuclear features must be performed. Furthermore, to classify an unknown sample using image analysis, a classification rule must be designed and evaluated. Unfortunately, statistical evaluation methods used in the literature of nuclear image analysis are often inappropriate. The present article discusses some of the difficulties in statistical evaluation of nuclear image analysis, and a study of cervical cancer is presented in order to illustrate the problems. In conclusion, some of the most severe errors in nuclear image analysis occur in analysis of a large feature set, including few patients, without confirming the results on an independent data set. To select features, Bonferroni correction for multiple test is recommended, together with a standard feature set selection method. Furthermore, we consider that the minimum requirement of performing statistical evaluation in nuclear image analysis is confirmation of the results on an independent data set. We suggest that a consensus of how to perform evaluation of diagnostic and prognostic features is necessary, in order to develop reliable tools for clinical use, based on nuclear image analysis.

Many are called, but few are chosen. Feature selection and error estimation in high dimensional spaces

We address the problems of feature selection and error estimation when the number of possible feature candidates is large and the number of training samples is limited. A Monte Carlo study has been performed to illustrate the problems when using stepwise feature selection and discriminant analysis. The simulations demonstrate that in order to find the correct features, the necessary ratio of number of training samples to feature candidates is not a constant. It depends on the number of feature candidates, training samples and the Mahalanobis distance between the classes. Moreover, the leave-one-out error estimate may be a highly biased error estimate when feature selection is performed on the same data as the error estimation. It may even indicate complete separation of the classes, while no real difference between the classes exists. However, if feature selection and leave-one-out error estimation are performed in one process, an unbiased error estimate is achieved, but with high variance. The holdout error estimate gives a reliable estimate with low variance, depending on the size of the test set.

On-Line Fat Content Classification of Inhomogeneous Pork Trimmings Using Multispectral near Infrared Interactance Imaging

A novel system for on-line measurement of fat content in inhomogeneous pork trimmings is presented. The system allows near infrared (NIR) energy to interact with the meat using non-contact optics while it is travelling in large plastic boxes on a conveyor belt. A comparison was made between the log of the inverse of the interactance NIR spectra [log(1/T)], standard normal variate (SNV) and extended multiplicative signal correction (EMSC) as techniques for the correction of physical light scattering due to colour and textural differences, height variation and temperature fluctuations, depending on whether the meat was warm-cut or cold-cut. EMSC gave the best prediction results; a robust partial least squares regression using two factors resulted in a root mean square error (RMSEP) of 1.9% on 20 kg batches of inhomogeneous meat trimmings. The model was fully tested twice in an on-line environment at a slaughter house and performed with a RMSEP of 3.4% for a fat range of 8–55% in the first industrial trial and 2.82% in the second industrial trial.

Automatic segmentation of overlapping fish using shape priors

We present results from a study where we segment fish in images captured within fish cages. The ultimate goal is to use this information to extract the weight distribution of the fish within the cages. Statistical shape knowledge is added to a Mumford-Shah functional defining the image energy. The fish shape is represented explicitly by a polygonal curve, and the energy minimization is done by gradient descent. The images represent many challenges with a highly cluttered background, inhomogeneous lighting and several overlapping objects. We obtain good segmentation results for silhouette-like images containing relatively few fish. In this case, the fish appear dark on a light background and the image energy is well behaved. In cases with more difficult lighting conditions the contours evolve slowly and often get trapped in local minima.

Texture Classification of Mouse Liver Cell Nuclei Using Invariant Moments of Consistent Regions

A new texture analysis approach is applied to the problem of classification of pathological states from electron microscopy images of mouse liver cell nuclei. For each pixel in the image, a region of consistent connected neighbouring pixels is extracted, forming a local texel of pixels belonging to the same gray level population. The geometric properties of each texel is described by invariant moment-based features. A recently developed method for fast and exact computation of Cartesian geometric moments is utilized. Each cell nucleus is characterized by a feature vector, giving the average feature values of both the bright and the dark structures. A leave-one-out classification is performed, using 4 different classes of cells (normal, proliferating, precancer and cancer).

A Multi-Element Detector System for Intelligent Imaging: I-ImaS

I-ImaS is a European project aiming to produce new, intelligent X-ray imaging systems using novel APS sensors to create optimal diagnostic images. Initial systems concentrate on mammography and encephalography. Later development will yield systems for other types of radiography such as industrial QA and homeland security. The I-ImaS system intelligence, due to APS technology and FPGAs, allows real-time analysis of data during image acquisition, giving the capability to build a truly adaptive imaging system with the potential to create images with maximum diagnostic information within given dose constraints. A companion paper deals with the DAQ system and preliminary characterization. This paper considers the laboratory X-ray characterization of the detector elements of the I-ImaS system. The characterization of the sensors when tiled to form a strip detector will be discussed, along with the appropriate correction techniques formulated to take into account the misalignments between individual sensors within the array. Preliminary results show that the detectors have sufficient performance to be used successfully in the initial mammographic and encephalographic I-ImaS systems under construction and this paper will further discuss the testing of these systems and the iterative processes used for intelligence upgrade in order to obtain the optimal algorithms and settings.

Adaptive image content-based exposure control for scanning applications in radiography

I-ImaS (Intelligent Imaging Sensors) is a European project which has designed and developed a new adaptive X-ray imaging system using on-line exposure control, to create locally optimized images. The I-ImaS system allows for real-time image analysis during acquisition, thus enabling real-time exposure adjustment. This adaptive imaging system has the potential of creating images with optimal information within a given dose constraint and to acquire optimally exposed images of objects with variable density during one scan. In this paper we present the control system and results from initial tests on mammographic and encephalographic images. Furthermore, algorithms for visualization of the resulting images, consisting of unevenly exposed image regions, are developed and tested. The preliminary results show that the same image quality can be achieved at 30-70% lower dose using the I-ImaS system compared to conventional mammography systems.

A scanning system for intelligent imaging: I-ImaS

I-ImaS (Intelligent Imaging Sensors) is a European project aiming to produce adaptive x-ray imaging systems using Monolithic Active Pixel Sensors (MAPS) to create optimal diagnostic images. Initial systems concentrate on mammography and cephalography. The on-chip intelligence available to MAPS technology will allow real-time analysis of data during image acquisition, giving the capability to build a truly adaptive imaging system with the potential to create images with maximum diagnostic information within given dose constraints. In our system, the exposure in each image region is optimized and the beam intensity is a function not only of tissue thickness and attenuation, but also of local physical and statistical parameters found in the image itself. Using a linear array of detectors with on-chip intelligence, the system will perform an on-line analysis of the image during the scan and then will optimize the X-ray intensity in order to obtain the maximum diagnostic information from the region of interest while minimizing exposure of less important, or simply less dense, regions. This paper summarizes the testing of the sensors and their electronics carried out using synchrotron radiation, x-ray sources and optical measurements. The sensors are tiled to form a 1.5D linear array. These have been characterised and appropriate correction techniques formulated to take into account misalignments between individual sensors. Full testing of the mammography and cephalography I-ImaS prototypes is now underway and the system intelligence is constantly being upgraded through iterative testing in order to obtain the optimal algorithms and settings.

Ultrastructural Texture Analysis as a Diagnostic Tool in Mouse Liver Carcinogenesis

Nuclear texture, which reflects the overall structure of the chromatin, may be used to detect early as well as later stages of malignancy. In this study, texture analysis was applied to four groups of liver cells in mice: normal and regenerating liver, hyperplastic nodules, and hepatocellular carcinomas. The best discriminating set of features was selected based on a training data set. The model was then tested on an independent series of 10 hyperplastic nodules and 6 hepatocellular carcinomas. A correct classification rate of 95% was obtained on the training data set and 100% accuracy was obtained on the test set. This kind of image analysis technique offers an opportunity to identify and describe the nuclear changes related to carcinogenesis, and the present results demonstrate the possible use of digital texture analysis as a diagnostic aid in tumor pathology.

Effects of Many Feature Candidates in Feature Selection and Classification

We address the problems of analyzing many feature candidates when performing feature selection and error estimation on a limited data set. A Monte Carlo study of multivariate normal distributed data has been performed to illustrate the problems. Two feature selection methods are tested: Plus-1-Minus-1 and Sequential Forward Floating Selection. The simulations demonstrate that in order to find the correct features, the number of features initially analyzed is an important factor, besides the number of samples. Moreover, the sufficient ratio of number of training samples to feature candidates is not a constant. It depends on the number of feature candidates, training samples and the Mahalanobis distance between the classes. The two feature selection methods analyzed gave the same result. Furthermore, the simulations demonstrate how the leave-one-out error estimate can be a highly biased error estimate when feature selection is performed on the same data as the error estimation. It may even indicate complete separation of the classes, while no real difference between the classes exists.