%Aigaion2 BibTeX export from Idiap Publications
%Thursday 21 November 2024 01:01:17 PM

@TECHREPORT{Raposo_Idiap-Com-01-2021,
         author = {Raposo, Geoffrey},
       keywords = {deep learning, generalization, Interpretability, transfer learning, Tuberculosis Detection},
       projects = {Idiap},
          month = {7},
          title = {Active tuberculosis detection from frontal chest X-ray images},
           type = {Idiap-Com},
         number = {Idiap-Com-01-2021},
           year = {2021},
    institution = {Idiap},
            url = {https://gitlab.idiap.ch/bob/bob.med.tb},
       abstract = {In this study, we investigate the benefits of automatic Pulmonary Tuberculosis (PTB) detection methods based on radiological signs found on CXR.  Contrary to direct scoring from images, implemented in most related work, indirect detection offers natural interpretability of automated reasoning.  We identify generalization difficulties for direct detection models trained exclusively on the modest amount of publicly available CXR images from PTB patients.  We subsequently show that a model, pre-trained on tens of thousands of CXR images using automatically annotated radiological signs, offers a more adequate base for development.  By relaying radiological signs through a simple linear classifier, one is able to obtain state-of-the-art results on three publicly available datasets (test AUC on Montgomery County-MC: 0.97, Shenzhen-CH: 0.90, and Indian-IN: 0.93).  We further discuss limitations imposed by the limited number of PTB-specific radiological signs available on public datasets, and evaluate possible performance gains that could be obtained if more were available (test AUC MC: 0.98, CH: 0.98, IN: 0.93).},
            pdf = {https://publications.idiap.ch/attachments/reports/2021/Raposo_Idiap-Com-01-2021.pdf}
}