Study On Algorithms For Melanoma Recognition Based On Dermoscopy

Melanoma’s morbidity and mortality keep growing in recent years, which seriously threaten human health. Once the melanoma cells metastasized, the current medical cure for the disease is powerless. The best treatment of melanoma is to resect tumor through early diagnosis. In general, the early diagnosis mainly depends on naked eye observation and pathological biopsy. However, tissue pathology biopsy may cause unnecessary trauma on patients, and observation with naked eyes is of great subjectivity. Therefore, the medical commuity is in urgent need for researches on computer-aided detection diagnosis.Aiming at the processing of the dermoscopy images with different form, color, size and noise features, and this thesis mainly conducts research on image noise removal, lesion region segmentation, feature extraction, recognition of skin lesions and its classification. The main contents of the thesis are:1. For dermoscopy images with diverse noise features, this paper has studied algorithms on noise removal and image segmentation. Noise removal includes man-made noises with black frames and bubbles and inherent noises with hair etc. The algorithms this paper put forward require little calculations and could well eliminate noise. The segmentation algorithm advocates that the fusion segmentation algorithm proposed based on Markov random field segmentation framework can improve the robustness of a single segmentation algorithm;2. This thesis finished the study on the extraction algorithm of dermoscopy image features, which extract low-level dermoscopy images features based on features designing method, including shape, color and texture, and senior dermoscopy images features based on features learning methods. In view of features learning methods, this paper put forward using sparse coding based on SIFT operators to show features and improving pooling algorithm to adapt to the regional structure characteristics of the lesions area. This paper also advances area pool algorithm;3. This thesis finished the study on the classification algorithm of dermoscopy image recognition. For the extracted low-level and senior level features, the author designed classifiers respectively. For a more reasonable and effective use of two kinds of features, the author fused them before classification and fused the results after classification, known as the co-training methods. Finally, this paper gets the final classification result using multi classifier voting mechanism;4. This thesis has evaluated the results of single classifier and fusing classifiers of two learning features based on the specificity, sensitivity and accuracy of the evaluation criterion and the receiver operating curve(ROC). Experimental results show that the results of fusion classifier receive a higher specificity, sensitivity and accuracy compared with the results of single classifier. Meanwhile, multiple sample recognitions result show that the robustness of fusion classifiers has improved.