{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/deconstructing-bias-on-skin-lesion-datasets","title":"(De)Constructing Bias on Skin Lesion Datasets","arxiv_id":"1904.08818","date":"2019-04-18","proceeding":null,"authors":["Alceu Bissoto","Michel Fornaciali","Eduardo Valle","Sandra Avila"],"abstract":"Melanoma is the deadliest form of skin cancer. Automated skin lesion analysis\nplays an important role for early detection. Nowadays, the ISIC Archive and the\nAtlas of Dermoscopy dataset are the most employed skin lesion sources to\nbenchmark deep-learning based tools. However, all datasets contain biases,\noften unintentional, due to how they were acquired and annotated. Those biases\ndistort the performance of machine-learning models, creating spurious\ncorrelations that the models can unfairly exploit, or, contrarily destroying\ncogent correlations that the models could learn. In this paper, we propose a\nset of experiments that reveal both types of biases, positive and negative, in\nexisting skin lesion datasets. Our results show that models can correctly\nclassify skin lesion images without clinically-meaningful information:\ndisturbingly, the machine-learning model learned over images where no\ninformation about the lesion remains, presents an accuracy above the AI\nbenchmark curated with dermatologists' performances. That strongly suggests\nspurious correlations guiding the models. We fed models with additional\nclinically meaningful information, which failed to improve the results even\nslightly, suggesting the destruction of cogent correlations. Our main findings\nraise awareness of the limitations of models trained and evaluated in small\ndatasets such as the ones we evaluated, and may suggest future guidelines for\nmodels intended for real-world deployment.","url_abs":"http://arxiv.org/abs/1904.08818v1","url_pdf":"http://arxiv.org/pdf/1904.08818v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"deconstructing-bias-on-skin-lesion-datasets","repo_url":"https://github.com/alceubissoto/deconstructing-bias-skin-lesion","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[{"task_slug":"machine-learning","task_name":"BIG-bench Machine Learning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1904.08818","atlas_url":"https://app.syntology.ai/?focus=1904.08818","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}