@article {Uegami2021.07.21.21260920, author = {Wataru Uegami and Andrey Bychkov and Mutsumi Ozasa and Kazuki Uehara and Kensuke Kataoka and Takeshi Johkoh and Yasuhiro Kondo and Hidenori Sakanashi and Junya Fukuoka}, title = {MIXTURE of human expertise and deep learning{\textemdash}Developing an explainable model for predicting pathological diagnosis and survival in patients with interstitial lung disease}, elocation-id = {2021.07.21.21260920}, year = {2021}, doi = {10.1101/2021.07.21.21260920}, publisher = {Cold Spring Harbor Laboratory Press}, abstract = {Interstitial pneumonia is a heterogeneous disease with a progressive course and poor prognosis, at times even worse than those in the main cancer types. Histopathological examination is crucial for its diagnosis and estimation of prognosis. However, the evaluation strongly depends on the experience of pathologists, and the reproducibility of diagnosis is low.Herein, we propose MIXTURE (huMan-In-the-loop eXplainable artificial intelligence Through the Use of REcurrent training), a method to develop deep learning models for extracting pathologically significant findings based on an expert pathologist{\textquoteright}s perspective with a small annotation effort. The procedure of MIXTURE consists of three steps as follows. First, we created feature extractors for tiles from whole slide images using self-supervised learning. The similar looking tiles were clustered based on the output features and then pathologists integrated the pathologically synonymous clusters. Using the integrated clusters as labeled data, deep learning models to classify the tiles into pathological findings were created by transfer-learning the feature extractors. We developed three models for different magnifications.Using these extracted findings, our model was able to predict the diagnosis of usual interstitial pneumonia, a finding suggestive of progressive disease, with high accuracy (AUC 0.90). This high accuracy could not be achieved without the integration of findings by pathologists. The patients predicted as UIP had significantly poorer prognosis (five-year overall survival [OS]: 55.4\%) than those predicted as non-UIP (OS: 95.2\%). The Cox proportional hazards model for each microscopic finding and prognosis pointed out dense fibrosis, fibroblastic foci, elastosis, and lymphocyte aggregation as independent risk factors. We suggest that MIXTURE may serve as a model approach to different diseases evaluated by medical imaging, including pathology and radiology, and be the prototype for artificial intelligence that can collaborate with humans.Competing Interest StatementThe authors have declared no competing interest.Funding StatementThis paper is based on results obtained from a project, JPNP20006,commissioned by the New Energy and Industrial Technology Development Organization (NEDO).Author DeclarationsI confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.YesThe details of the IRB/oversight body that provided approval or exemption for the research described are given below:Ethics Committee of Nagasaki University Hospital (protocol 19012107).All necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived.YesI understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).YesI have followed all appropriate research reporting guidelines and uploaded the relevant EQUATOR Network research reporting checklist(s) and other pertinent material as supplementary files, if applicable.YesData to be provided upon a reasonable request.}, URL = {https://www.medrxiv.org/content/early/2021/07/23/2021.07.21.21260920}, eprint = {https://www.medrxiv.org/content/early/2021/07/23/2021.07.21.21260920.full.pdf}, journal = {medRxiv} }