Quantifying Detection Quality in Presence of Adversarial Inputs in Dermatological Images

• Item type: SIG Technical Reports(1)
• 公開日: 2021-02-25
• タイトル: Quantifying Detection Quality in Presence of Adversarial Inputs in Dermatological Images
• タイトル / 言語: en
• タイトル: Quantifying Detection Quality in Presence of Adversarial Inputs in Dermatological Images
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: セッション3-2
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Department of Information & Communication Engineering, The University of Tokyo
• 著者所属: exMedio Inc.
• 著者所属: Department of Information & Communication Engineering, The University of Tokyo
• 著者所属(英): en
• 著者所属(英): Department of Information & Communication Engineering, The University of Tokyo
• 著者所属(英): en
• 著者所属(英): exMedio Inc.
• 著者所属(英): en
• 著者所属(英): Department of Information & Communication Engineering, The University of Tokyo
• 著者名: Sourav, Mishra; Hideaki, Imaizumi; Toshihiko, Yamasaki
• 著者名(英): Sourav, Mishra; Hideaki, Imaizumi; Toshihiko, Yamasaki
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: We have tested deep learning based detection on dermatological conditions commonly encountered in clinical settings. Despite successes in diagnosing critical and morbid conditions such as Melanoma, it is not well understood if such models can reduce the patient burden on doctors by screening benign diseases. Most projects traditionally use pristine data acquired in controlled conditions. This may not reflect regular clinical workflows where image quality is non-ideal. We test the performance of deep learning methods on such data by simulating imperfections on user-submitted images of common disease labels. In our study, we have found the overall predictions change significantly despite robust training, contraindicating the maturity to enter mainstream medical diagnostics.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: We have tested deep learning based detection on dermatological conditions commonly encountered in clinical settings. Despite successes in diagnosing critical and morbid conditions such as Melanoma, it is not well understood if such models can reduce the patient burden on doctors by screening benign diseases. Most projects traditionally use pristine data acquired in controlled conditions. This may not reflect regular clinical workflows where image quality is non-ideal. We test the performance of deep learning methods on such data by simulating imperfections on user-submitted images of common disease labels. In our study, we have found the overall predictions change significantly despite robust training, contraindicating the maturity to enter mainstream medical diagnostics.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AA11131797
• 書誌情報: 研究報告コンピュータビジョンとイメージメディア（CVIM）; 巻 2021-CVIM-225, 号 24, p. 1-6, 発行日 2021-02-25
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8701
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会