Hello!! Our research for detecting skin neoplasms was published in PLOS Medicine (https://doi.org/10.1371/journal.pmed.1003381). It is a large scale detection study, and we discussed various biases that may exist in the ML research.
We performed an external validation using all single lesion biopsied cases for the past 10 years. A total of 10,426 cases, consisting of about 40,000 clinical photographs were used for the retrospective cohort validation. Almost all kinds of skin tumors (43 neoplasms) were included in the analysis.
In the diagnosis of skin neoplasms, our algorithm showed comparable performance in the reader test, but the performance was inferior to that of dermatologists in a tertiary care center. The lack of data relevancy may bring the inferiority of the algorithm. Various biases that lead to the overestimation of algorithms in the medical field were discussed in the paper.
Since 2018, real-time web DEMOs have been available at http://modelderm.com (Classification Model; Model Dermatology) and http://rcnn.modelderm.com (Detection Model; Model Dermatology with RCNN), to demonstrate reproducibility. The development of our model was previously published in JAMA Dermatology in 2019 (https://jamanetwork.com/journals/jamadermatology/article-abstract/2756346).
Recently, we made a derivative application (http://ai.skindx.net) for assisting laypeople to select the appropriate type of skin clinic.
We opened the peer-review history. The questions and responses in the process of peer-review are available, which may be also interesting for the researchers.
Thank you.
Submitted January 17, 2021 at 06:44AM by whria78 https://ift.tt/35Ms42F via TikTokTikk
No comments:
Post a Comment