Invasive malignant melanoma has risen in incidence for decades. Dermoscopy aids in detection of the earliest and most curable stage of melanoma. The long lead times required for appointments with specialists trained in dermoscopy and a lack of dermoscopy training for mid-level practitioners, often the first to encounter a melanoma, contribute to a delay in melanoma detection at the most curable stage. Recent advances in classifier architectures and computing power have enabled development of advanced image-processing tools to improve diagnostic accuracy, exceeding that of dermatologists. Extensive collections of dermoscopy images are available to enable better training for deep learning computational techniques. However, automated methods still need to provide sufficient accuracy for reliable melanoma screening for in-person and virtual consultations. Computational techniques are less successful for limited data sets, such as those of specific features. Our hypothesis is: Annotation of critical skin lesion structures for small-scale data sets by multiple experts enables automatic structure detection that can improve automatic lesion discrimination ability. To investigate this hypothesis, the proposed three-year project is a collaboration with the largest publicly available skin lesion image archive to create a platform that promotes international participation in dataset annotation, curation, and validation to facilitate advances in skin lesion analysis research. This platform will allow global experts to annotate and validate dermoscopic skin lesion image datasets to identify and segment critical structures, providing an archive integrated into the publicly accessible International Skin Imaging Collaborative (ISIC) image collection. Skin lesion feature analysis and validation studies will be conducted to identify critical features contributing to improved melanoma discrimination from benign mimics. The proposed research will investigate the fusion of existing deep learning techniques based on the entire skin lesion with deep learning techniques trained using localized annotation information for guided deep learning of specific skin lesion structures for feature extraction and diagnosis. We will perform statistical analyses to determine relevant annotated skin lesion small-scale feature set sizes and the significance of feature detection and diagnostic rates between standard deep learning techniques, targeted localized feature set-based guided deep learning training, and fused (combined) results for hypothesis assessment. Hypothesis-driven feature detection and diagnosis results from this study will impact small-scale dataset analysis for deep learning algorithm development. As part of the proposed project, seminars will be provided to pre-college students through STEM programs such as Project Lead the Way. Further educational and training opportunities will be provided to undergraduate students pursuing pre-med, biological sciences, data scienc...