A hierarchical three-step superpixels and deep learning framework for skin lesion classification.

Skin cancer is one of the most common and dangerous cancer that exists worldwide. Malignant melanoma is one of the most dangerous skin cancer types has a high mortality rate. An estimated 196,060 melanoma cases will be diagnosed in 2020 in the USA. Many computerized techniques are presented in the past to diagnose skin lesions, but they are still failing to achieve significant accuracy. To improve the existing accuracy, we proposed a hierarchical framework based on two-dimensional superpixels and deep learning. First, we enhance the contrast of original dermoscopy images by fusing local and global enhanced images. The entire enhanced images are utilized in the next step to segmentation skin lesions using three-step superpixel lesion segmentation. The segmented lesions are mapped over the whole enhanced dermoscopy images and obtained only segmented color images. Then, a deep learning model (ResNet-50) is applied to these mapped images and learned features through transfer learning. The extracted features are further optimized using an improved grasshopper optimization algorithm, which is later classified through the Naïve Bayes classifier. The proposed hierarchical method has been evaluated on three datasets (Ph2, ISBI2016, and HAM1000), consisting of three, two, and seven skin cancer classes. On these datasets, our method achieved an accuracy of 95.40%, 91.1%, and 85.50%, respectively. The results show that this method can be helpful for the classification of skin cancer with improved accuracy.

Automated multi-level pathology identification techniques for abnormal retinal images using artificial neural networks.

AIM: To automatically classify abnormal retinal images from four different categories using artificial neural networks with a high degree of accuracy in minimal time to assist the ophthalmologist in subsequent treatment planning. METHODS: We used 420 abnormal retinal images from four different categories (non-proliferative diabetic retinopathy, central retinal vein occlusion, central serous retinopathy and central neo-vascularisation membrane). Green channel extraction, histogram equalisation and median filtering were used as image pre-processing techniques, followed by texture-based feature extraction. The application of Kohonen neural networks for pathology identification was also explored. RESULTS: The approach described yielded an average classification accuracy of 97.7% with ±0.8% deviation for individual categories. The average sensitivity and the specificity values are 96% and 98%, respectively. The time taken by the Kohonen neural network to achieve these accurate results was 300±40 s for the 420 images. CONCLUSION: This study suggests that the approach described can act as a diagnostic tool for retinal disease identification. Simultaneous multi-level classification of abnormal images is possible with high accuracy using artificial neural networks. The results also suggest that the approach is time-efficient, which is essential for ophthalmologic applications.