Researchers have developed a deep learning-based artificial intelligence (AI) algorithm that improves the accuracy of diagnosing acute intracranial haemorrhage (AIH) on brain CT scans. The algorithm combines haemorrhage detection with anomaly detection, using a large dataset to achieve optimal results. The study was conducted by a team of researchers from several institutions, including Seoul National University Hospital, South Korea, and the University of Edinburgh, UK.
AIH is a life-threatening condition that requires prompt diagnosis, as a delayed diagnosis can lead to devastating consequences. However, diagnosing AIH using brain CT scans remains a challenge for physicians, as false negatives may delay the correct diagnosis, and false positives will lead to unnecessary examinations. Additionally, the high volume of imaging data that requires assessment places a significant burden on radiologists who need to maintain diagnostic accuracy and efficiency.
The new AI algorithm was trained using a large dataset and achieved an overall area under the receiver operating characteristic (AUROC) of 0.992 and 0.977 for patient-wise and slice-wise analyses, respectively. The patient and slice-wise analyses indicated a sensitivity of 94.4% and 79.0%, and a specificity of 98.2% and 99.3%, respectively. The algorithm improves the diagnostic performance of clinicians of varying expertise levels, as demonstrated in a retrospective multi-reader study.
The researchers used a joint recurrent neural network (CNN-RNN) approach for haemorrhage detection and unsupervised training for anomaly detection. This approach overcomes the limitations of the supervised haemorrhage detection process used in conventional AI algorithms, leading to an improvement in diagnostic performance.
The study demonstrates the potential of deep learning-based AI algorithms for assisting radiologists and physicians in their clinical diagnosis workflow, increasing diagnostic accuracy and enabling prompt diagnosis and improved management of various conditions. The researchers hope that this new AI algorithm can be used to improve the diagnosis and management of AIH, potentially saving lives and improving patient outcomes.
Link to study: https://www.nature.com/articles/s41746-023-00798-8