Gokhan Bilgin

Computerized classification of intraductal breast lesions using histopathological images

In the diagnosis of preinvasive breast cancer, some of the intraductal proliferations pose a special challenge. The continuum of intraductal breast lesions includes the usual ductal hyperplasia (UDH), atypical ductal hyperplasia (ADH), and ductal carcinoma in situ (DCIS). The current standard of care is to perform percutaneous needle biopsies for diagnosis of palpable and image-detected breast abnormalities. UDH is considered benign and patients diagnosed UDH undergo routine follow-up, whereas ADH and DCIS are considered actionable and patients diagnosed with these two subtypes get additional surgical procedures. About 250 000 new cases of intraductal breast lesions are diagnosed every year. A conservative estimate would suggest that at least 50% of these patients are needlessly undergoing unnecessary surgeries. Thus, improvement in the diagnostic reproducibility and accuracy is critically important for effective clinical management of these patients. In this study, a prototype system for automatically classifying breast microscopic tissues to distinguish between UDH and actionable subtypes (ADH and DCIS) is introduced. This system automatically evaluates digitized slides of tissues for certain cytological criteria and classifies the tissues based on the quantitative features derived from the images. The system is trained using a total of 327 regions of interest (ROIs) collected across 62 patient cases and tested with a sequestered set of 149 ROIs collected across 33 patient cases. An overall accuracy of 87.9% is achieved on the entire test data. The test accuracy of 84.6% is obtained with borderline cases (26 of the 33 test cases) only, when compared against the diagnostic accuracies of nine pathologists on the same set (81.2% average), indicates that the system is highly competitive with the expert pathologists as a stand-alone diagnostic tool and has a great potential in improving diagnostic accuracy and reproducibility when used as a “second reader” in conjunction with the pathologists.

Segmentation of Hyperspectral Images via Subtractive Clustering and Cluster Validation Using One-Class Support Vector Machines

This paper presents an unsupervised hyperspectral image segmentation with a new subtractive-clustering-based similarity segmentation and a novel cluster validation method using one-class support vector (SV) machine (OC-SVM). An estimation of the correct number of clusters is an important task in hyperspectral image segmentation. The proposed cluster validity measure is based on the power of spectral discrimination (PWSD) measure and utilizes the advantage of the inherited cluster contour definition feature of OC-SVM. Hence, this novel cluster validity method is referred to as SV-PWSD. SVs found by OC-SVM are located at the minimum distance to the hyperplane in the feature space and at the arbitrarily shaped cluster contours in the input space. SV-PWSD guides the segmentation/clustering process to find the optimal number of clusters in hyperspectral data. Because of the high computational load of subtractive clustering and OC-SVM, a subset of the image (only ground-truth data) is initially used in the clustering and validation phases. Then, it is proposed to use K-nearest neighbor classification, with the already clustered subset being used as training data, to project the initial clustering results onto the entire data set.

Unsupervised classification of hyperspectral-image data using fuzzy approaches that spatially exploit membership relations

This letter presents unsupervised hyperspectral-image classification based on fuzzy-clustering algorithms that spatially exploit membership relations. Not only is the conventional fuzzy c-means approach used to demonstrate the advantage of using membership relations but also Gustafson-Kessel clustering, which uses an adaptive distance norm, is, for the first time, used for the segmentation of hyperspectral images. A novel approach to include spatial information in the segmentation process is achieved by making use of spatial relations of fuzzy-membership functions among neighbor pixels. Two- and three-dimensional Gaussian filtering of fuzzy-membership degrees is utilized for this purpose. A novel phase-correlation-based similarity measure is used to further enhance the performance of the proposed approach by taking spatial relations into account for pixels with similar spectral characteristics only. It is shown that the proposed approach provides superior clustering performance for hyperspectral images.

Cell segmentation in histopathological images with deep learning algorithms by utilizing spatial relationships

In many computerized methods for cell detection, segmentation, and classification in digital histopathology that have recently emerged, the task of cell segmentation remains a chief problem for image processing in designing computer-aided diagnosis (CAD) systems. In research and diagnostic studies on cancer, pathologists can use CAD systems as second readers to analyze high-resolution histopathological images. Since cell detection and segmentation are critical for cancer grade assessments, cellular and extracellular structures should primarily be extracted from histopathological images. In response, we sought to identify a useful cell segmentation approach with histopathological images that uses not only prominent deep learning algorithms (i.e., convolutional neural networks, stacked autoencoders, and deep belief networks), but also spatial relationships, information of which is critical for achieving better cell segmentation results. To that end, we collected cellular and extracellular samples from histopathological images by windowing in small patches with various sizes. In experiments, the segmentation accuracies of the methods used improved as the window sizes increased due to the addition of local spatial and contextual information. Once we compared the effects of training sample size and influence of window size, results revealed that the deep learning algorithms, especially convolutional neural networks and partly stacked autoencoders, performed better than conventional methods in cell segmentation.

Kernel Fukunaga–Koontz Transform Subspaces for Classification of Hyperspectral Images With Small Sample Sizes

In this letter, a novel supervised classification approach is presented for the classification of hyperspectral images using kernel Fukunaga-Koontz transform (KFKT). The Fukunaga-Koontz transform (FKT) is originally a powerful target detection method used in remote sensing tasks, and it is an especially good classification tool for two-class problems. The traditional FKT method has been kernelized for increasing the nonlinear discrimination ability and capturing higher order of statistics of data. The proposed approach in this letter aims to solve the multiclass problem by regarding one class as target that is tried to be separated from the remaining classes (as clutter) like one-against-all methodology. The KFKT provides superior performance in the classification of hyperspectral data even using small number of samples because of nonlinear separability of data in higher dimensional space. The experiments confirm that KFKT has better and promising results than FKT and support vector machine in classification of hyperspectral images.

Classification of histopathological images using convolutional neural network

In this work, classification of cellular structures in the high resolutional histopathological images and the discrimination of cellular and non-cellular structures have been investigated. The cell classification is a very exhaustive and time-consuming process for pathologists in medicine. The development of digital imaging in histopathology has enabled the generation of reasonable and effective solutions to this problem. Morever, the classification of digital data provides easier analysis of cell structures in histopathological data. Convolutional neural network (CNN), constituting the main theme of this study, has been proposed with different spatial window sizes in RGB color spaces. Hence, to improve the accuracies of classification results obtained by supervised learning methods, spatial information must also be considered. So, spatial dependencies of cell and non-cell pixels can be evaluated within different pixel neighborhoods in this study. In the experiments, the CNN performs superior than other pixel classification methods including SVM and k-Nearest Neighbour (k-NN). At the end of this paper, several possible directions for future research are also proposed.

Enhancing land use classification with fusing dual-polarized TerraSAR-X and multispectral RapidEye data

Abstract. The contribution of dual-polarized synthetic aperture radar (SAR) to optical data for the accuracy of land use classification is investigated. For this purpose, different image fusion algorithms are implemented to achieve spatially improved images while preserving the spectral information. To compare the performance of the fusion techniques, both the microwave X-band dual-polarized TerraSAR-X data and the multispectral (MS) optical image RapidEye data are used. Our test site, Gediz Basin, covers both agricultural fields and artificial structures. Before the classification phase, four data fusion approaches: (1) adjustable SAR-MS fusion, (2) Ehlers fusion, (3) high-pass filtering, and (4) Bayesian data fusion are applied. The quality of the fused images was evaluated with statistical analyses. In this respect, several methods are performed for quality assessments. Then the classification performances of the fused images are also investigated using the support vector machines as a kernel-based method, the random forests as an ensemble learning method, the fundamental k-nearest neighbor, and the maximum likelihood classifier methods comparatively. Experiments provide promising results for the fusion of dual polarimetric SAR data and optical data in land use/cover mapping.

Embolic Doppler ultrasound signal detection via fractional Fourier transform

Computerized analysis of Doppler ultrasound signals can aid early detection of asymptomatic circulating emboli. For analysis, physicians use informative features extracted from Doppler ultrasound signals. Time -frequency analysis methods are useful tools to exploit the transient like signals such as Embolic signals. Detection of discriminative features would be the first step toward automated analysis of embolic Doppler ultrasound signals. The most problematic part of setting up emboli detection system is to differentiate embolic signals from confusing similar wave-like patterns such as Doppler speckle and artifacts caused by tissue movement, probe tapping, speaking etc. In this study, discrete version of fractional Fourier transform is presented as a solution in the detection of emboli in digitized Doppler ultrasound signals. An accurate set of parameters are extracted using short time Fourier transform and fractional Fourier transform and the results are compared to reveal detection quality. Experimental results prove the efficiency of fractional Fourier transform in which discriminative features becomes more evident.

Segmentation of histopathological images with Convolutional Neural Networks using Fourier features

The study aims to boost the success of the segmentation results by evaluating spatial relations in the segmentation of histopathalogical images. In the first step Fourier features are extracted from RGB color space of digital histopathalogical images. Training data sets are formed by selecting equal number of different cellular and extra-cellular structures in spatial domain from the images. Classification models of each training data set is obtained by utilizing Convolutional Neural Network (CNN), Support Vector Machine (SVM) and k-Nearest Neighbor (k-NN) methods. Visual and numerical outputs which are obtained from supervised training methods are presented for comparison purpose in the experimental results section.

A comparative study on manifold learning of hyperspectral data for land cover classification

This paper focuses on the land cover classification problem by employing a number of manifold learning algorithms in the feature extraction phase, then by running single and ensemble of classifiers in the modeling phase. Manifolds are learned on training samples selected randomly within available data, while the transformation of the remaining test samples is realized for linear and nonlinear methods via the learnt mappings and a radial-basis function neural network based interpolation method, respectively. The classification accuracies of the original data and the embedded manifolds are investigated with several classifiers. Experimental results on a 200-band hyperspectral image indicated that support vector machine was the best classifier for most of the methods, being nearly as accurate as the best classification rate of the original data. Furthermore, our modified version of random subspace classifier could even outperform the classification accuracy of the original data for local Fisher’s discriminant analysis method despite of a considerable decrease in the extrinsic dimension.