Abstract:
The recent emergence of machine learning and deep learning methods for medical image analysis has enabled the development of intelligent medical imaging-based diagnosis systems that can assist physicians in making better decisions about a patient's health. In particular, skin imaging is a field where these new methods can be applied with a high rate of success. Skin cancer diseases bring silent death due to ability of spreading to other parts of the body. Small variations in lesion pattern can cause mixing of non-melanoma with melanoma lesions for a layman. Early detection of these diseases can reduce the level of disease severity. Segmentation is the pre-requisite process in most of the computer aided diagnosis systems for medical imaging. Skin lesion segmentation has been in research since years but the traditional techniques for lesion segmentation result in poor segmentation when applied to diseased images and produces false positives in the presence of lesions. Presence of different artifacts makes segmentation of skin lesion very difficult. Abnormal growth of artifacts can appear as false positives and can degrade the performance of the diagnosis systems. It can be avoided only when false structures are removed while extracting the lesion. Within this framework, automated skin lesion segmentation is proposed which achieves high accuracy segmentation of skin lesion. To address this issue, this thesis proposes deep leaning for skin lesion segmentation. Our proposed architecture is 31 layers deep with same filter size. The validity of the proposed techniques is tested on two publically available databases of PH2 and ISIC 2017. Experimental results show the efficiency of the proposed approaches. The proposed method gives Dice Coefficient of 92.3% Accuracy of 95.01 %, Sensitivity of 93.44% and Specificity of 95.6 % for PH2 Dataset while Dice Coefficient of 85.5% Accuracy of 93.6%, Sensitivity of 82.9and Specificity of 95.1% for ISIC 2017 Dataset.
