Conjunctival melanoma detection using deep learning in smartphone images

The approach of our study is anterior segment examination using a smartphone image with deep learning system, which is similar to previous dermatology literature of skin cancer detection. This study was based on publicly accessible ocular images on the web. We collected anterior segment images from the Google image search engine. The strategy of image search was based on the following keywords: “conjunctiva”, “pterygium”, “conjunctival nevus”, “conjunctival melanoma”, and “conjunctival malignant melanoma”. The most recent image from the search was performed on Autumn 31, 2020. Images were manually classified by two board-certified ophthalmologists, and the ambiguous images were excluded to clarify the image domains.

本研究采用智能手机拍摄图像结合深度学习系统（deep learning system）开展眼前节（anterior segment）检查，该研究思路与既往皮肤癌检测领域的皮肤病学相关文献相仿。本研究依托网络公开可获取的眼部图像完成，数据源自谷歌图片搜索引擎（Google image search engine）抓取的眼前节图像。本次图像检索使用的关键词包括：结膜（conjunctiva）、翼状胬肉（pterygium）、结膜痣（conjunctival nevus）、结膜黑色素瘤（conjunctival melanoma）以及恶性结膜黑色素瘤（conjunctival malignant melanoma）。本次检索获取的最新图像采集于2020年秋季31日。所有图像均由两名认证眼科医师（board-certified ophthalmologists）完成人工分类，并剔除存在歧义的图像以明确图像类别范畴。

搜集汇总

背景与挑战

背景概述

该数据集旨在通过深度学习系统检测智能手机拍摄的眼前段图像中的结膜黑色素瘤，基于2020年秋季前从网络公开收集的图像，由眼科专家手动分类以明确图像领域，借鉴了皮肤病癌检测的研究方法。