Colour-based computer image processing approach to melanoma diagnosis

© 2017 Dr. Sahar Sabbaghi MahmoueiMelanoma is one of the most prevalent skin cancers in the world. The incidence and mortality rates of melanoma in Australian populations have been sharply increasing over the last decades. For instance, it is represented that two in three Australian develops some form of skin cancer before they reach the age of 70. Most melanoma can be cured if diagnosed and treated in the early stages. Over the past decades, advances in dermoscopy technology has made it an effective technique used in early diagnosis of malignant melanoma. Dermoscopy allows the clinicians to visualise different colours and examine microstructures in the skin that are not visible to the naked eye. This clear view of the skin reduces screening errors and improves the diagnostic accuracy of pigmented skin lesions significantly. However, it has been demonstrated that the performance and accuracy of melanoma diagnosis using dermoscopic images manually depend on the quality of the image and the clinical experience of the dermatologists. Several medical diagnosis methods have been developed to help dermatologists interpret the structures revealed through dermoscopy, such as the pattern analysis, the ABCD rule, the 7-point checklist, the Menzies method, CASH algorithm, the Chaos and Clues algorithm and the BLINCK algorithm. However, the diagnosis criteria used in assessing the potential of melanoma may be easily overlooked in early melanomas, or be misinterpreted as a benign mole, mainly attending to the subjectivity of clinical interpretation. Also, human judgement is often hardly reproducible. Therefore, clinical diagnosis is still challenging, especially with equivocal pigmented lesions, which leading to the accuracy of melanoma diagnosis by expert dermatologists remains at 75–84%. Only biopsy or excision of a pigmented skin lesion can provide a definitive diagnosis. However, a biopsy can rise metastasizing, in addition to be being invasive and an unpleasant experience to the patient. Therefore, to minimise the diagnostic errors, and provide a reliable second independent opinion to dermatologists, the development of computerised image analysis techniques is of paramount importance. In the last decade, several computer-aided diagnosis (CAD) systems have been proposed to tackle this problem. However, the diversity of existing problems makes any further contributions greatly appreciated. Moreover, it is widely acknowledged that much higher accuracy is required for computer-based system to be considered reliable and trustworthy enough by clinicians, therefore be adopted routinely in their diagnostic process. With the aim of improving some of existing approaches and developing new techniques to facilitate accurate, fast and more reliable computer-based diagnosis of melanoma, this thesis describes novel image processing approaches for computer-aided detection on selected subset of medical criteria that play an important role in the diagnosis of melanoma. This ensures that the features used by the system have a medical meaning, making it possible for the dermatologist to understand and validate the automated diagnosis. One of the contributions of this thesis is to develop a fast and accurate colour detection method. It is observed that colours may vary slightly in dermoscopy images, because of different levels of contrast. This may lead to difficulty in the perception of colours by dermatologists, resulting in subjectivity of clinical diagnosis. A computer-assisted system for quantitative colour identification is highly desirable for dermatologists to use. However, these colour variations within the lesion makes colour detection a challenging process. To tackle this challenge, a comprehensive colour detection procedure is conducted in this thesis. It incorporates a colour enhancement step to overcome the problems of poor contrast. Since colours perceived by the human observer are produced by a mixture of pixel values, we performed a summarised representation of colours by subdividing the colour space into colour clusters, using QuadTree clustering, comprising a set of RGB values. The proposed method employed a colour palette, to mimic human interpretation of of lesion colours in determining the type and the number of colours in melanocytic lesion images. In addition, a set of parameters such as colour feature set, texture feature set, and locational features is extracted to numerically describe the colour properties of each segmented block throughout the lesion. Furthermore, when comparing colour distribution in malignant melanomas (MMs) and benign melanomas (BMs), a significant difference in the number of colours in the two populations is detected. Also, the proposed method shown that the type of colour can greatly affect in the diagnosis outcome. The effectiveness of the proposed colour detection system is evaluated by comparing the obtained results with those obtained by using expert dermatologists. The highest correlation coefficients for detecting the type of colour is observed for red and blue–grey, which, in respect of the image set used in this thesis, signifies the most important colours for diagnosis purposes. The overall performance of the proposed system is evaluated by using machine learning techniques, and the best classification results, AUC of 0.93, are achieved by using kernel SVM classifier. Another contribution of this thesis is to provide meaningful visualisation of streak, and extract features to determine the relative importance of streak in classifying the skin lesion into two class of benign and malignant. To find streaks, a trainable B-COSFIRE filter applied in dermoscopy images to detect a prototype pattern of interest (bar-shaped structures) such as streak. Its application consists of convolution with Difference of Gaussian (DoG) filters, its blurring responses; shifting the blurred responses and estimate a point-wise weighted Geometric Mean (GM). To also account the different thickness and structure of streak a bank of B-COSFIRE filter is applied on the image with different orientation and rotation. Then to identify valid streaks from candidate streak lines, clinical criteria such as number of streaks in the images and the orientation pattern analysis is calculated and the false detected lines are removed. The result includes line segments that indicate the pixels that belong to streaks are displayed. Also, a set of features derived from streaks (such as geometrics, colour and texture features) are fed to three different classifiers for classifying images. We achieved an accuracy of 93.3% for classifying dermoscopy images into benign and malignant on 807 dermoscopy images. Furthermore, a novel, comprehensive and highly effective application of deep learning (stacked sparse auto-encoders) is examined in this thesis for classification of skin lesion. The model learns a hierarchal high-level feature representation of skin image in an unsupervised manner. The stacked sparse auto-encoder discovers latent information features in input images (pixel intensities). These high-level features are subsequently fed into a classifier for classifying dermoscopy images. In addition, we proposed a new deep neural network architecture based on bag-of-features (BoF) model, which learns high-level image representation and maps images into BoF space. We have shown that using BoF as the input to the auto-encoder can easily improve the performance of neural network in comparison with the raw input images. The proposed method is evaluated on a test set of 244 skin images and result shown that the deep BoF model achieves higher classification scores (with SE = 95.4% and SP = 94.9%) in compare to the raw input images. Our contributions will improve automated diagnosis of melanoma using dermoscopy images