Open AccessNeuralmed trIA: Automated Screening of Urgent Findings on Head CT Scans
Author(s) -
André Coutinho Castilla,
Priscilla Koch Wagner,
Jessicados Santos de Oliveira,
Maria Fernanda Barbosa Wanderley,
Anthony Moreno Eigier
Publication year - 2021
Language(s) - English
Resource type - Conference proceedings
DOI - 10.5327/1516-3180.342
Subject(s) - normalization (sociology) , artificial intelligence , artificial neural network , medicine , computer science , predictive value , categorical variable , cross entropy , pattern recognition (psychology) , machine learning , sociology , anthropology
Background:NeuralMed is a startup specialized in the development of artificial intelligence applications in medicine. Objectives: We developed a screening system for head computed tomography scans (HCT) that prioritizes patients with urgent pathologies. This study describes the development of the algorithm for detection and localization of intracranial hemorrhage. Design and setting: This is an observational study on HCT performed at hospitals in State of São Paulo. Methodology: The algorithm was built from 8432 HCT. We used unenhanced axial images after post-processing and normalization. The set is split into 80% for training, 10% for validation and 10% for testing. We used a MobileNet network pre-trained with ImageNet weights combined with a long short-term memory with categorical cross-entropy as loss function. The model’s outputs are hemorrhage, no findings, and other pathologies. A gradient class activation map was applied to identify and localize the hemorrhages. Results: Internal validation showed an area under the ROC curve (AUROC) of 96%, sensitivity of 87% and positive predictive value (PPV) of 96%. External validation was performed on 125 exams collected in a period after the training group obtaining an AUROC of 86%, sensitivity of 78% and PPV of 81%. Conclusion: After detecting an identifying the bleedings it is possible to order the patient queue prioritizing those most likely to have abnormalities and life-threatening situations. The algorithm also indicates the lesion location by showing the regions of the images that most activated the neural network.