Urgent diagnosis and triage in Intracranial Haemorrages with Machine Learning

Keywords: intracranial haemmorage, artificial intelligence, neurosurgery, deep learning


Acute intracranial hemorrhages, regardless of their type, are pathologies with high mortality and require rapid diagnosis and treatment, however the patient group who will benefit most from early operation is operated later than the patient group with nless favorable outcome, because they do not admit with a severe clinical presentation. In this study, we aimed to evaluate a deep learning model that can distinguish the presence of intracranial hemorrhage in a small data set. Material Method: 3 healthy patients and 5 patients with intracranial hemorrhages were randomly seleceted for the study from the qure.ai Cranial CT database. The data set was created with a total of 200 CT cross-section images, 100 of which were hemorrhagic and 100 were healthy, and it was divided into three groups as training, validation and test set. The artificial neural network was trained in the training set and its accuracy was tested in the validation set, the accuracy did not improve after reaching around 80% and the training of the artificial neural network was stopped. Later, this artificial neural network was evaluated in the test set. Results: The deep learning model was run on the test set. Results were as follows; Sensitivity 90.0%, Specificity: 70.0%, Positive Predictive Value: 75.0%, Negative Predictive Value: 87.5% Total Accuracy: 80.0%. The deep learning model made only one false-negative assessment in 20 crosss-sections that it had never seen before. As a result, we think that a deep learning model can produce highly accurate results even if they are trained in a small data set and potentially be used for rapid triage in emergency departments.  
How to Cite
B. Şiyar, “Urgent diagnosis and triage in Intracranial Haemorrages with Machine Learning”, js, vol. 2, no. 2, pp. 115-120, Dec. 2020.