Automatic detection of pituitary microadenoma from magnetic resonance imaging using deep learning algorithms

Abstract

The risks of misdiagnosed pituitary microadenoma is high. We designed a convolutional neural network (CNN) based computer-aided diagnosis (CAD) system to retrospectively diagnose patients with pituitary microadenoma. A total 5,540 pituitary magnetic resonance (MR) images from 1,108 participants were recruited. MRI images were randomly stratified into non-overlapping cohorts (training set, validation set and test set) to establish five different CNN models. The best CNN model is the ResNet with a diagnostic accuracy of 94%, which outperforms the diagnosis accuracy of our radiologists (64%-85%). The accuracy of our CAD system is further confirmed in additional MR datasets. The diagnostic accuracy of our ResNet model is comparable to the proficiency of a radiologist with 5-10 years’ experience. In summary, this is the first report showing that the CAD system is a viable tool for diagnosing pituitary microadenoma. CAD system is applicable to radiology departments, especially in primary health care institutions.