Dermatology Skin Lesion Image Analysis

Dermatology skin lesion image analysis research has been ongoing at Missouri S&T (previously UMR) since the 1980s. Our group has been successful in finding over 20 key dermoscopic structures in melanoma, melanoma mimics, and nonmelanoma skin cancers using iterative structure-based analysis. Structures are chosen to reduce system errors which are concentrated in a few classes: amelanotic/featureless, regressed, and small in situ melanomas, and the most difficult benign lesions to identify: Clark nevi, lentigines, and seborrheic keratoses [1]. Preliminary research detecting and using annotated lesion structures [2-7] and lesion artifacts [8,9] guided by clinical experience [10] with image processing and deep learning-based architectures for segmentation and lesion classification. Structures that computer vision can detect allow more accurate diagnosis of challenging lesions. We have used inspectable structure segmentation deep learning outputs with associated lesion classification to facilitate the correlation of structures with misdiagnosed lesions; this gives a path to exploring the relationship between expert annotated structures with automated segmentation and lesion classification.

Our current research focuses on multi-expert annotation of key skin lesion structures from small-scale datasets with computational learning and fusion with existing computational whole-lesion methods to enhance automated structure detection and lesion discrimination capability. Data validation is an important component of identifying key structures in skin lesions due to inter-expert variability. Deep learning techniques have, up to now, been applied to whole dermoscopy images. Processing at the whole-lesion (global) scale may miss critical structures. Resulting errors may be overcome by detecting critical structures in a lesion. Curating small-scale datasets of annotated key structures associated with specific lesion types by multiple experts provides two significant, novel contributions: 1) validation of key structures used for lesion diagnosis that are disseminated in the research community and 2) fusion of expert and machine learning in a reinforcement learning framework to promote enhanced lesion classification and interpretability of machine learning classification decisions.

The International Skin Imaging Collaboration (ISIC) archive offers the largest publicly available collection of dermoscopic images for skin lesion analysis. However, the ISIC 2019 dataset suffers from several acquisition artifacts such as hair occlusions, vignetting, surgical ink markings, and ruler scales, all of which compromise its utility in deep learning workflows by introducing non-biological visual structures. A rigorously annotated artifact archive of ISIC 2019 is given here that includes binary masks for key artifact types, including ink marks, vignetting, hair and ruler marks. The annotations were obtained through manual (ink marks and vignetting) and semi-automatic (hairs and ruler marks) systematic review process. Separate .zip files with the annotation masks for the different artifact types are provided for the ISIC 2019 dataset. These artifact archive annotation collections provide an initial step to develop a multi-expert data annotation process with quality control that can be extended to dataset annotation of other critical skin lesion features for expert-in-the-loop, deep learning whole-lesion algorithm development, and data fusion for enhanced skin lesion analysis.