Content-based Image Retrieval System for clinical diagnosis of Pigmented Skin Lesions

The screening process of skin cancer and the capacity of storing the digital images in recent years are rapidly increasing at an alarming rate. These digital image contains a lot of useful diagnostic information, which are not efficiently accessed and used. This requires a way to quickly and accurately find an access to these images, also known as content-based image retrieval (CBIR) system. The CBIR systems are developed based on text-based and visual features-based search techniques. A few content-based image retrieval (CBIR) systems were developed in the past to search pigmented skin lesions (PSLs) based on visual features from a set of dermoscopy images. Those CBIR systems were limited to specific categories of PSLs and employed noneffective visual features. In this paper, an improved CBIR (Derma-CBIR) system using deep learning algorithms for PSLs is proposed by defining effective visual color and texture features for retrieving skin lesions. The recall, precision and rank statistical metrics are utilized to test and compare the performance of CBIR systems. The Derma-CBIR system is tested on a dataset of total 240 lesions (20 images per category) achieved an average recall of 0.921, precision of 0.875 and rank of 0.081. The obtained results indicate that the Derma-CBIR is effective when compared with other state-of-the-art CBIR systems. As a result, it can be used to assist clinical experts for maintaining PSLs images.