Optimal knowledge extraction for robust DELM automatic data classification in skin melanocytic lesions

The operational performance of a new MDASS (Melanoma Diagnosis Active Support System) prototype, able to distil optimal knowledge from acquired DELM (Digital EpiLuminescence Microscopy) data to automatically capture and reliably discriminate and quantify the stage of skin Malignant Melanoma (MM) disease evolution, is presented. The past decade has seen DELM techniques introduced in modern Dermatological departmental applications for general operational support. In a few cases DELM introduction was matched with the joint development of a Computer Assisted Digital Dermatology (CADD) computational environment. CADD systems may allow for rapid evaluation of initial reference and feature guidelines for dermatological lesion computer-automated classification feasibility and design projects, in particular for MM supported diagnosis, like MDASS (Melanoma Diagnosis Active Support System). In fact, Skin MM tumour is particularly dangerous for humans. Its malignant evolution lasts about 5 or 6 years and usually ends with subject death. Early diagnosis is a powerful mean of preventing this adverse evolution allowing sudden intervention, increasing the probability of subject recovery and survival. Unfortunately, a reduced number of well-trained dermatologic early diagnosis experts is available only. MDASS could be a key component for economic and reliable mass screening clinical plan and telediagnosis applications for a first discrimination level between malignant or benign skin melanocytic lesions. MDASS prototype is based on GEOGINE© (GEOmetrical enGINE) OMG (Ontological Model Generator) [14] computational kernel for geometric and colour geometry discrimination and classification to achieve robust, high reliable performance. As a reference feature selection, the work by Stanganelli and Kenet was taken into account. A careful Biomedical Engineering analysis supplied initial structured parameter template architecture, enabling to extract maximal information content from experimental dermatological data quickly. Automated classification dermatoscopical parameters can be divided into two main classes: Size Descriptor (point size, local, and global) and Intrinsic Descriptor (morphological, geometrical, chromatic, others). They offer the best structured operative synthetic description compromise, when compared to traditional specialistic ones. At present time, Structured Parameter Descriptors seem to represent the most suitable tool to elicit knowledge discovery for the effective development of reliable Active Support Systems in dermatoscopical diagnostic applications. MDASS prototype main results show disease classification procedure with distillation of minimal reference grids for pathological cases and ultimately achieve effective early diagnosis of melanocytic lesion and its follow-up. Furthermore, MDASS computational complexity and error analysis are discussed. Finally, system results are validated by carefully designed experiments with certified clinical reference database.