This paper examines the role of deep learning applied in medical setting, notably with an attempt to detect; by highlighting regions of interest of intracranial haemorrhages for the purpose of diagnosis and binary classification. Haemorrhagic strokes are a leading mortality cause affecting 13.7 million people with 5.5 million deaths annually. Data was accumulated from public and academic database source Kaggle, no human participants were identified nor used and the scans collected are owned by a third-party. To carry out the solution with regulated momentum a waterfall methodology was adapted for the target of completing the artefact. Results were gathered by training the deep-learning classifier on firstly, ~200 samples on 200 epoch which yield wrong class predictions. Secondly, a larger dataset was used ~7000 samples each split to 0.7 training and 0.3 testing split sets which achieved high accuracy, precision, and recall metrics at 20 epochs. Lastly, same samples were used at lower epoch ~5, to introduce some bias and increase justification argument and transparency of the model, on average the metrics deemed more realistic with value set at 0.98 recall and 0.99 everywhere else. To conclude, artefact successfully justified the aim for this paper but also unravels unanswered questions in adapting existing machine models for different problem settings. It is insistent to keep researching this field as deep learning posses’ issues that need to be addressed.
Janusz Snieg (25085325) , May 2023 University of Lincoln School of Computer Science