Hierarchical agglomerative clustering-based skin lesion detection with region based neural networks classification

Abstract

The skin cancer disease is widely spreading in the world due to abnormal increment of radiation. The segmentation of images is a crucial yet challenging aspect of the image processing process. It has emerged as a central focus in the study of image comprehension in recent years. Skin cancer is one of the critical diseases, which caused millions of deaths over the globe. The computer aided detection (CAD) of skin cancer can save the human lives. Further, the CAD models are resulted in poor segmentation and classification performance due to basic image processing, and machine learning models. Therefore, this work focused on development of skin lesion detection and classification (SLDC) using deep learning models. Initially, Guided image filtering (GIF) based preprocessing method, which performs noise removal, skin lesion enhancements, and basic hair removal operations. Then, skin lesion disease effected region is identified by using hierarchical agglomerative clustering (HAC). Then, multi-level features are extracted from segmented images. Finally, the region based neural network with long short term memory (RNN-LSTM) are used to perform the benign and malignant classes classification using trained features. The simulations conducted on International Symposium on Biomedical Imaging (ISBI) dataset shows that, the proposed RNN-SLDC system resulted in superior subjective and objective performance as compared to other approaches. The proposed method achieved 99.71% of segmentation accuracy, and 99.46% of classification accuracy.