Lina Choridah

Identification of malignant masses on digital mammogram images based on texture feature and correlation based feature selection

The most popular techniques in early breast cancer detection is using digital mammogram. However, the challenge lies in early and accurate detection the irregular masses with spiculated margin as the most common abnormality. This paper proposes an image classifier to classify the mammogram images. The abnormality that can be founded in mammogram image is classified into malignant, benign and normal cases. By applying Computer Aided Diagnosis (CAD), totally 12 features comprising of histogram and GLCM as the texture based features are extracted from the mammogram image. Correlation based feature selection (CFS) is used in this paper which reduces 50% of the features. Multilayer perceptron algorithm is applied to mammography classification by using these selected features. The experimental result shows that 40 digital mammograms data taken from private Oncology Clinic Kotabaru Yogyakarta was achieved 91.66% of accuracy. The approach can be beneficial to radiologists for more accurate diagnosis.

Analysis of digital mammograms for detection of breast cancer

Digital mammogram has become the most effective technique for early breast cancer detection. The most common abnormality that may indicate breast cancer is masses. The challenge lies in early and accurate detection to overcome the development of breast cancer that affects more and more women throughout the world. Computer Aided Diagnosis (CAD) is used to help the radiologist in interpretation and recognition the pattern of the mammogram abnormality. The main objective of this research is to perform and analyze the contrast enhancement and feature selection method in order to build a CAD to discriminate normal, benign, and malignant. Preprocessing needs to enhance the poor quality of image and remove the artifact caused by preprocessing step. ROI as the suspicious area segmented, and then extracted by texture feature approach. High dimensionality of feature is selected by feature selection technique and would be classified according to their class each other. The digital mammogram images are taken from the Private database of Oncology Clinic Kotabaru Yogyakarta. The dataset consists of 40 mammogram images with 14 benign cases, 6 malignant cases, and 20 normal cases. The proposed method in preprocessing step made the image enhanced and proved by MSE and PSNR value. Histogram and gray level co-occurrence matrix (GLCM) as the texture feature are used to extract the suspicious area. Correlation based feature selection (CFS) is used to select the best feature among 12 extracted features before. Mean, standard deviation, smoothness, angular second moment (ASM), entropy, and correlation are the best feature that guarantee the improvement of classification with less feature dimension. The result shows that the proposed method was achieved the accuracy 96.66%, sensitivity 96.73%, specificity 97.35% and ROC 96.6% It is expected to contribute for helping the radiologist as material consideration in decision-making.

Comparison between Automatic and Semiautomatic Thresholding Method for Mammographic Density Classification

Mammographic density is a novel independent risk factor of breast cancer that reflects the amount of fibroglandular tissue. Breast Imaging Reporting and Data System (BIRADS) density is one of the mammographic density classification schemes which are most widely used by radiologists. Initially, the method used for assessing mammographic density was subjective and qualitative. Recently however, the measurement of mammographic density is more objective and quantitative. In this paper, we propose an alternative model of breast cancer risk factor assessment based on a quantitative approach of density mammogram. This model consists of pre-processing, breast area counting, fibroglandular tissue area counting that uses maximum entropy and multilevel thresholds, and finally breast density counting to determine the risk factor of breast cancer. The proposed model has been tested on a private database from Oncology Clinic Kotabaru, Yogyakarta, Indonesia consisting of 30 mammograms and has been analyzed by some radiologists using the semiautomatic threshold. The result shows that percentage of mammographic density counted by maximum entropy threshold method has the accuracy, sensitivity and specificity of about 87%, 73% and 91% respectively compared to the semiautomatic thresholding method. On the other hand, the accuracy, sensitivity and specificity resulted from using multilevel threshold is about 93%, 87% and 96% respectively. The obtained results suggest that multilevel threshold is perfectly suited for getting quantitative measurement of mammographic density as one of the strongest risk factors for breast cancer.

Performance analysis of filtering techniques for speckle reduction on breast ultrasound images

Breast cancer is one of the most frequent types of cancer occurs in women and also the leading cause of womens death in the world. Ultrasonography (USG) is the common modality which is used as the diagnosis tool due to its low cost and portability. It is used to early detect breast cancer. In the USG, the degradation of image edge character is indistinct or speckle noise frequently occurred at the time of the acquisition. It is important to carry out an effective process to reduce speckle noise. Some filters are proposed for reducing the speckle. However, the results cannot preserve the edges and the details of the USG images. This paper aims to study the performance of some filtering techniques in reducing speckle noise that can preserve the edges and the details of breast USG images. Multiple parameter measurements to determine image quality among others mean square error (MSE), signal to noise ratio (SNR), speckle index (C), average difference (AD) and contrast to speckle ratio (CSR) are used to compare and evaluate of Wiener filter, anisotropic diffusion filter (ADF), Perona-Malik filter, detail preserved anisotropic diffusion (DPAD) filter, speckle reduction anisotropic diffusion (SRAD), anisotropic diffusion filter with memory based on speckle statistics (ADMSS) and maximum homogeneity over a pixel neighborhood (m-homog). Based on the measurement parameter can specify filter that are providing a robust filter for de-speckling, preserving edges and details, while improving contrast of breast ultrasound images.

Automated detection of breast cancer lesions using adaptive thresholding and morphological operation

One of the imaging modalities for early detection of breast cancer is ultrasonography (USG). The detection is based on lesions identification. Radiologists still manually conduct early detection of the lesions. Hence, the detection results tend to be subjective and may cause different interpretations due to the different level of knowledge and experience of the radiologists. In this research, the proposed method for the detection of automatic lesions uses 30 images consisting of 20 images benign lesions and 10 images of malignant lesions. At the pre-processing stage, adaptive median filter is applied and is followed by adaptive thresholding method at the segmentation process. The final stage uses morphological operations. The result shows that the proposed method successfully achieved the accuracy of 95.19%, sensitivity of 84.13% and specificity of 96.2%. These results indicate that the system can be used to assist the experts or operators from radiology team more objectively in breast cancer lession detection.

Performance analysis of image segmentation for breast ultrasound images

Breast cancer is one of the main causes of women mortality worldwide. Ultrasonography (USG) is other modalities than mammography that capable to support radiologists in diagnosing breast cancer. However, the diagnosis may come with different interpretation depending on the radiologists experience. Therefore, Computer-Aided Diagnosis (CAD) is developed as a tool for radiologists second opinion. CAD is built based on digital image processing of ultrasound (US) images which consists of several stages. Lesion segmentation is an important step in CAD system because it contains many important features for classification process related to lesion characteristics. This study provides a performance analysis and comparison of image segmentation for breast USG images. In this paper, several methods are presented such as a comprehensive comparison of adaptive thresholding, fuzzy C-Means (FCM), Fast Global Minimization for Active Contour (FGMAC) and Active Contours Without Edges (ACWE). The performance of these methods are evaluated with evaluation metrics Dice coefficient, Jaccard coefficient, FPR, FNR, Hausdorff distance, PSNR and MSSD parameters. Morphological operation is able to increase the performance of each segmentation methods. Overall, ACWE with morphological operation gives the best performance compare to the other methods with the similarity level of more than 90%.

Feature extraction for lesion margin characteristic classification from CT Scan lungs image

Lung cancer is one of the common cancer which occurred in both male and female. Revealed by WHO data, in 2012, this disease become one of the major cause of death in worldwide with the mortality rate about 1.59 million. An early detection of lung cancer by using Computed Tomography (CT) Scan can provide more opportunity to survive. However, the diagnosis of lung cancer by reading the CT scan image which performed by radiologists may lead to an error. A computer-based digital image processing is a solution to improve the accuracy and consistency in reading the CT Scan image result. This study aim is to identify the morphological characteristic of regular and irregular margins by using feature extraction method. In this research, image processing divided into several stages refer to the segmentation process with Otsu method, feature extraction with number of features such as convexity, solidity, circularity, and compactness, and the last is classification by using Multi Layer Perceptron (MLP). The classification process of features convexity, solidity, circularity, and compactness, resulted in the accuracy value of 85%, sensitivity of 85%, and specificity of 85%.

Classification of breast ultrasound images based on posterior feature

Ultrasonography (USG) check-up is a common way for breast cancer screening, but the result is highly subjective on the operator. Therefore, a system capable to objectively diagnose breast cancer is necessary. One of the features of breast cancer is posterior acoustic patterns. It categorized into four classes which are enhancement, shadowing, combined pattern, and no posterior acoustic feature. This paper proposes a scheme by extracting area suspected to have posterior acoustic features and background features. The dataset consists of 98 breast USG images which are classified into 69 posterior acoustic enhancement cases and 29 no posterior acoustic cases. Firstly, a pre-processing of breast USG images is conducted to eliminate speckle noise, marker, and label. Secondly, segmentation is using region growing method, and followed by extracting posterior area and its background. Feature extraction is conducted on both of areas using histogram method. Finally, classification is using Multilayer Perceptron (MLP). Performance of the proposed method successfully achieves accuracy of 87.79%, sensitivity of 92.75% and specificity of 82.75% using six histogram features. It shows that this method is succesful in classifying the breast USG images. Therefore, it has potential to be implemented in an automated breast computer aided diagnosis (CAD) system.

Improvement of Sample Selection: A Cascade-Based Approach for Lesion Automatic Detection

Computer-Aided Detection (CADe) system has a significant role as a preventative effort in the early detection of breast cancer. There are some phases in developing the pattern recognition on the CADe system, including the availability of a large number of data, feature extraction, selection and use of features, and the selection of the appropriate classification method. Haar cascade classifier has been successfully developed to detect the faces in the multimedia image automatically and quickly. The success of the face detection system must not be separated from the availability of the training data in the large numbers. However, it is not easy to implement on a medical image because of some reasons, including its low quality, the very little gray-value differences, and the limited number of the patches for the examples of the positive data. Therefore, this research proposes an algorithm to overcome the limitation of the number of patches on the region of interest to detect whether the lesion exists or not on the mammogram images based on the Haar cascade classifier. This research uses the mammogram and ultrasonography images from the breast imaging of 60 probands and patients in the Clinic of Oncology, Yogyakarta. The testing of the CADe system is done by comparing the reading result of that system with the mammography reading result validated with the reading of the ultrasonography image by the Radiologist. The testing result of the k-fold cross validation demonstrates that the use of the algorithm for the multiplication of intersection rectangle may improve the system performance with accuracy, sensitivity, and specificity of 76%, 89%, and 63%, respectively.

A Comparative Study of Thresholding Algorithms on Breast Area and Fibroglandular Tissue

One of the independently risk factors of breast cancer is mammographic density reflecting the composition of the fibroglandular tissue in breast area. Tumor in the mammogram is precisely complicated to detect as it is covered by the density (the masking effect). The determination of mammographic density may be implemented by calculating percentage of mammographic density (quantitative and objective approaches). Thereby, the use of a proper thresholding algorithm is highly required in order to obtain the fibroglandular tissue area and breast area. The mammograms used in the research were derived from Oncology Clinic, Yogyakarta that had been verified by Radiologists using semi-automatic thresholding. This research was aimed to compare the performance of the thresholding algorithm using three parameters, namely: PME, RAE and MHD. Zack Algorithm had the best performance to obtain the breast area with PME, RAE and MHD of about 0.33%, 0.71% and 0.01 respectively. Meanwhile, there were two algorithms having good performance to obtain the fibroglandular tissue area, i.e. multilevel thresholding and maximum entropy with the value for PME (13.34%; 11:27%), RAE (53.34%; 51.26%) and MHD (1:47; 33.92) respectively. The obtained results suggest that zack algorithm is perfectly suited for getting breast area than multilevel thresholding and maximum entropy for getting fibroglandular tissue. It is one of the components to determine risk factors of breast cancer based on percentage of breast density. Keywords: Thresholding Algorithm; Breast Area; fibroglandular Area