Computer Vision News - August 2016

Every month, Computer Vision News reviews a challenge related to our field. If you can’t find time to read challenges, but are interested in the new methods proposed by the scientific community to solve them, this section is for you. This month we have chosen to review the Skin Lesion Analysis toward Melanoma Detection held at ISBI 2016 , the International Symposium on Biomedical Imaging and hosted by ISIC, the International Skin Imaging Collaboration. The paper describing design and implementation of the challenge is here . Background Skin cancer is a major public health concern, with over 5M newly diagnosed cases in the U.S. each year. Among them, melanoma is one of the most lethal forms of skin cancer, expected to be cause over 10K deaths in the U.S. in 2016. Though melanomas are often first visually recognized by patients, many potentially curable melanomas are not detected until advanced stages. Dermoscopy , an imaging technique that eliminates the surface reflection of skin, has been introduced to improve diagnostic performance and reduce melanoma mortality. The goal of the ISIC effort is to support development of automated melanoma diagnosis algorithms from dermoscopic images. Challenge The challenge was divided into three parts corresponding to each stage of lesion analysis: lesion segmentation (participants were asked to submit automated predictions of lesion segmentations from dermoscopic images in the form of binary masks); lesion dermoscopic feature detection (participants were asked to automatically detect two clinically defined dermoscopic features, ”globules” and ”streaks”, while pattern detection involved both localization and classification - see figure below); and lesion classification (participants were asked to classify images as either being benign or malignant). 79 submissions from a group of 38 participants make it the largest standardized and comparative study to date for melanoma diagnosis in dermoscopic images. They were compared using common metrics of accuracy, sensitivity and specificity . Findings show that segmentation methods appear to be within the range to provide utility for annotation of additional data, though further analysis is needed before concluding that automated techniques are statistically equivalent to expert annotation. Challenge Example lesion dermoscopic pattern annotations. Left column: original images. Center column: extracted SLIC superpixels. Right column: Positive superpixel annotations highlighted, overlayed over original image. Multiple colors correspond to multiple human annotators. Top row: example for “Globule” annotation label. Bottom row: example for “Streak” annotation label [image from article b y Gutman, Codella, Celebi, Helba, Marchetti, Mishra, Halpern] Skin Lesion Analysis for Melanoma Detection Computer Vision News Challenge 21