Optimized Veterinary Thermographic Image Classification using Support Vector Machines and Noise Mitigation

摘要

Introduction: Thermography has gained popularity in both human and veterinary medicines in recent years. Unlike other traditional imaging methods, it is non-invasive, fast, and inexpensive and it produces harmless radiation. Many researchers have combined thermography and pattern classification to create an automatic diagnosis tool, which is a computer software capable of detecting pathologies using the thermographic images. In medical thermography, breast cancer detection is most highly researched topic. Various pattern classification algorithms have performed excellently well in thermographic image classification problems; few names are a k-nearest neighbor, artificial neural network, fuzzy inference system, and random forest classifiers. In the field of pattern recognition, support vector machine is one of the most popular supervised learning algorithm. In many pattern classification problems, it has performed superiorly well compared to the other algorithms. The support vector machines as pattern classifiers can be combined to thermography to develop a powerful diagnostic tool to detect various pathologies.;Objectives: This study investigates the support vector machines as the potential classifiers in veterinary thermographic image classification to detect the canine bone cancer disease, canine anterior cruciate ligament rupture, and feline hyperthyroid disease. In addition, the possibilities of gray level quantization with 32 gray levels and 3x3 averaging filter as the noise mitigation techniques have been studied.;Experimental Methods: 166 thermographic images, including both normal and abnormal cases, of three pathologies are used in this study. The thermographs are further divided into four experimental sets corresponding to pathology and body part: bone cancer, elbow knee, bone cancer, wrist, anterior cruciate ligament, and feline hyperthyroidism (shaved). The research study has been conducted in two stages. In the first stage, the gray scale thermographic images with 256 gray levels are used to extract the gray level co-occurrence matrix texture features. Instead of selecting one texture distance, the features are extracted for four different values of texture distance - 1, 3, 7 and 9. The extracted feature data are used to train and test the support vector machine (SVM) classifiers using leave-one-out cross-validation. In the second stage, two noise mitigation techniques, namely gray-level quantization with 32 levels and 3x3 average filters are applied to mitigate the noise pattern due to hair and Gaussian noise respectively. Again, the feature extraction and pattern classification experiments are performed using the noise mitigated images.;Results: The best classification rates achieved for canine bone cancer in elbow/knee part are the accuracy of 92.68%, the sensitivity of 85% and the specificity of 100%. Similarly, the thermographs of canine bone cancer in wrist body part can be classified with an accuracy of 96.65%, a sensitivity of 100%, and a specificity of 92.86%. Also, the very high accuracy is achieved for anterior cruciate ligament rupture; the best classification results are 96.3% accuracy, 100% sensitivity, and 95% specificity. Among the three pathologies, the feline hyperthyroidism has the best classification result with an accuracy of 100%, a sensitivity of 100%, and a specificity of 100.00%. For all pathologies, the texture distance of 7 and 9 works the best. With noise mitigation using 3x3 average filter and 32-gray level quantization, the performances of the support vector machine classifiers are not improved.;Conclusion: These experimental results indicate that GLCM texture features and SVM has the potential to classify veterinary thermographic images to detect canine bone cancer, canine anterior cruciate ligament rupture, and feline hyperthyroidism. And, the gray level quantization with 32 gray levels and 3x3 average filter are not useful methods to mitigate the noise in thermographs to improve the thermographic image classification when support vector machines and GLCM texture features are used.