Automated Segmentation of Skin Lesion from Dermoscopic Images (T-0700) (MFN 4236)

Abstract

Skin cancer is one of the most prevalent types of cancer in our world. Diagnosis of skin cancer needs specialized equipment, doctors and continuous monitoring. Patients living in remote areas normally cannot access such facilities. To overcome these barriers of access, Computer Aided Diagnostics, an emerging field in computer science, often called telemedicine, is being considered a promising approach. Image processing for Computer Aided Diagnostics has three key steps, i.e. Segmentation, Feature Extraction and Classification. And segmentation is a critical step to ensure the accuracy of the system. Segmentation is required to separate the affected skin area from the healthy part of the skin. The resultant segmented image contains important characteristics that are inputs into the next step of feature extraction in the overall sequence of automated analysis. In this research, segmentation of skin cancer images was performed on dermatoscope images by using wavelet-based techniques along with other morphological operations. It has been found that, in some cases, wavelet transformations provide better results as compared to other techniques like thresholding, clustring, region growing, watershedding and GVF snakes. One obvious obvervation is of removing external artifacts like gel, water bubbles and dark hair around the surface affected by cancer, i.e. these artifacts are removed with less effort. Experiments also showed that images with Blue channel from RGB are better segmented as compared to other grayscale conversion techniques. Different wavelets were also tested and it was found that the ’bior6.8’ Cohen- Daubechies-Feauveau biorthogonal wavelet provides better results as compared to other wavelets like Haar. Overall, we achieved more 95.3% True Detection Rate(TDR) with 6.5% False Positive Rate(FPR) on 200 dermatoscope images on PH2 dataset, which has not been test before on 200 image with wavelet implementation, and using other techniques in combination with wavelets is expected to yield higher accuracy.