Anjan Kumar Ghosh

Breast abnormality detection through statistical feature analysis using infrared thermograms

The utilization of medical infrared thermography in breast abnormality detection is mostly due to its radiation-free, non-invasive and painless nature. Infrared breast thermography is an alternative breast imaging modality that can detect those tumors or early changes which are undetectable by the gold standard method X-ray mammography. However, breast cancer is a highly treatable disease, with 97% chances of survival if getting detected earlier. Thus, early detection of breast cancer using infrared breast thermography may improve the survival rate of breast cancer patients. The temperature pattern in both breasts of a healthy breast thermogram is closely symmetrical. Hence, a small asymmetry in the temperature pattern of the left and right breast may signify a breast abnormality. There are a series of texture features that play a vital role in asymmetry analysis of breast thermograms. This paper mainly emphasizes on investigating those statistical features, which can adequately differentiate the healthy breast thermograms from pathological breast thermograms. A survey work on texture features used by various authors for asymmetry detection is provided in this work. Our analysis is performed on 30 healthy and 30 abnormal breast thermograms of existing DMR (Database for Mastology Research) Database. The analysis and experimental results show that among the first order statistical features, the mean difference, skewness, entropy and standard deviation are the most efficient features that contribute most towards the asymmetry detection.

An Approach for Automatic Pain Detection through Facial Expression

Abstract Automatic pain detection is an emerging area of investigation with convenient applications in health care. The variation in facial expression often provides a clue for occurrence of pain. It provides an important window for the person who cannot verbally describe or rate their level of pain. To meet up the specific necessities, a framework has been designed for extraction of features from the face for automatic pain detection through facial expression. In this framework, Gabor filtering and Principal Component Analysis (PCA) are used as contributive steps that improves the performance of the system in terms of accuracy. To verify the accuracy and robustness of the system, experiments have been conducted on UNBC-McMaster Shoulder Pain Expression Archive Database at both frame level (person dependent) and image level (person independent). The methodology achieves 87.23% accuracy for detection of pain at frame level. Also the methodology achieves 82.43% accuracy for classifying the frames between four pain level (i.e. PSPI of 0, 1, 2 and >=3). The success rate of the methodology for pain detection at image level is 95.5%.

A Study and Analysis of Hybrid Intelligent Techniques for Breast Cancer Detection Using Breast Thermograms

The growing incidence and mortality rate of breast cancer draw the attention of the researchers to develop a technique for improving the survival rate of the cancer patients. Medical infrared thermography (MIT) with sensitivity 90 % has proved itself as a safe and promising method for early breast cancer detection. Moreover, an abnormal breast thermogram can signify breast pathology. The accurate classification and diagnosis of these breast thermograms is one of the major problem in decision making for treatments, which leads to the utilization of hybrid intelligent system in breast thermogram classification. Hybrid intelligent system plays a vital role in survival prediction of a breast cancer patient, and it is highly significant in decision making for treatments and medications. The primary objective of a hybrid intelligent system is to take the advantages of its constituent models and at the same time lessen their limitations. This chapter is an attempt to highlight the reliability of infrared breast thermography and hybrid intelligent system in breast cancer detection and diagnosis. A detailed overview of infrared breast thermography including its principles and role in early breast cancer detection is described here. Several research works are carried out by various researchers to identify the breast pathology from breast thermograms by using hybrid intelligent techniques which include extraction and analysis of several statistical features. A study of research works related to feature extraction and classification of breast thermograms using various types of hybrid classifiers is also included in this chapter.

Analysis and Performance Evaluation of ICA-Based Architectures for Face Recognition

Prediction of the best ICA architecture for face recognition systems is somewhat complicated. This paper shows how the recognition performance of both architectures depends on the nature of feature vectors rather than several criteria such as different databases, number of subjects, and number of principle components. The investigation finds that Architecture-II yields the better performance than Architecture-I based on face feature vectors. The experiments are done on different face datasets like FERET, ORL, CVL, and YALE.

Discriminative feature selection for breast abnormality detection and accurate classification of thermograms

Infrared breast thermography with the potential of predicting the future risk of developing breast cancer, has been considered as an early breast abnormality detection tool. This paper investigates the importance of selecting the discriminative features for improving the classification accuracy of the infrared thermography based breast abnormality detection systems. Mann-Whitney-Wilcoxon statistical test has been used here to select the best discriminative features from a feature set of 24 features, extracted from each breast thermogram of DBT-TU-JU and DMR databases. Three set of features: FStat, STex and SSigFS generated from these 24 extracted features are then fed into six most widely used classifiers for comparing the efficiency of each feature set in breast abnormality detection. The experimental results show that among all three feature sets, statistically significant feature set (SSigFS) provides more accuracy in discriminating the abnormal breast thermograms from the normal.

Contrast Restoration of Fog-Degraded Image Sequences

Poor visibility in the presence of fog is a major problem for many applications of computer vision. Still image and video systems are typically of limited use in poor visibility condition as the degraded images/frames lack visual vividness and offer low visibility of the scene contents. This paper investigates the defogging effects on images and frames by using a fast defogging method on our own newly developed database, namely Society of Applied Microwave Electronics Engineering and Research-Tripura University (SAMEER-TU) database which consists of 5,390 color images and 10 videos captured in foggy as well as in clear condition. The first step of the method ensures contrast enhancement yielding better global visibility, but the images/frames containing very dense fog still suffer from low visibility. In that case, Luminance and chromatic weight map have been used. Finally for verifying the robustness of the method, qualitative assessment evaluation in respect of peak-signal-to-noise ratio (PSNR) and root-mean-square error (RMSE) is introduced as a contributory step in this paper.

Singular value based characterization and analysis of thermal patches for early breast abnormality detection

The purpose of this study is to develop a novel breast abnormality detection system by utilizing the potential of infrared breast thermography (IBT) in early breast abnormality detection. Since the temperature distributions are different in normal and abnormal thermograms and hot thermal patches are visible in abnormal thermograms, the abnormal thermograms possess more complex information than the normal thermograms. Here, the proposed method exploits the presence of hot thermal patches and vascular changes by using the power law transformation for pre-processing and singular value decomposition to characterize the thermal patches. The extracted singular values are found to be statistically significant (p < 0.001) in breast abnormality detection. The discriminability of the singular values is evaluated by using seven different classifiers incorporating tenfold cross-validations, where the thermograms of the Department of Biotechnology-Tripura University-Jadavpur University (DBT-TU-JU) and Database of Mastology Research (DMR) databases are used. In DMR database, the highest classification accuracy of 98.00% with the area under the ROC curve (AUC) of 0.9862 is achieved with the support vector machine using polynomial kernel. The same for the DBT-TU-JU database is 92.50% with AUC of 0.9680 using the same classifier. The comparison of the proposed method with the other reported methods concludes that the proposed method outperforms the other existing methods as well as other traditional feature sets used in IBT based breast abnormality detection. Moreover, by using Rank1 and Rank2 singular values, a breast abnormality grading (BAG) index has also been developed for grading the thermograms based on their degree of abnormality.

Bruise detection in apples using infrared imaging

Defects in apples cause food safety concerns touching the general public and strongly affect the commodity market. Due to the increasing incidence, the detection of bruises is a challenge now a days especially when the bruises are not visible externally. Infrared imaging provides an important window for detection of bruises that are not visible externally. The study has been investigated on the infrared images of both fresh and bruised apples. For our investigation, a new dataset has been designed by maintaining standard acquisition protocol to improve the potentiality and accuracy of the thermograms. The contribution of the investigation includes asymmetric analysis of the acquired thermograms using first and second order statistical analysis and fractal analysis of pre-processed thermograms. Automatic detection of the bruised region was done using anisotropic diffusion and K-means clustering for investigation of the spread of the bruised region.

Visibility enhancement techniques for fog degraded images: A comparative analysis with performance evaluation

Low visibility is regarded as the fundamental cause for increasing number of accidents. When bad weather condition exists mainly due to fog, haze, snow, darkness, etc., the driver is unable to observe a distinct view of route. Out of the bad atmospheric condition, fog is one of the major sources of the accident because the visibility of fog remains very low which is less than 1 km. It is the natural phenomenon that decreases the contrast and color fidelity of objects in the captured image and makes the object difficult to see through naked eyes. The main goal of this paper is to perform a comparative analysis of some well-known visibility enhancement techniques. This paper also implemented three well-known fog removal algorithms, and for assessing the efficiency of the methods used, qualitative assessment evaluation is accomplished along with comparative statistical analysis and algorithms efficiency comparison.

Automated edge detection of breast masses on mammograms

Edge of a breast mass is one of the indicators of breast abnormality detection. In a mammogram, round and circumscribed masses indicate benign changes and malignant masses usually has speculated (irregular) boundary. The paper has encountered a fundamental problem of active contour model which was first proposed by Kass et al. The problem encountered here is generation of initial contour points manually selected by users. Thus the positions of initial contour points will vary with human perspective, which is very difficult to identify actual and accurate contour points. To overcome this problem to some extent, sobel edge detection method is used as a prior step of active contour model. Experiments have been tested on a dataset of 160 mammograms collected from Mini-MIAS benchmark database and compared with sobel edge detection method. In experiments, 92.5% segmentation accuracy has been obtained with sensitivity 93% and 85% specificity where the sobel edge detection method shown very less segmentation accuracy of 84% with 91% sensitivity and 50% specificity. Time complexity and detection error have been also analysed for proposed method, ideal high pass filter, sobel edge detection, hough transform and active contour model.