Members of our team have contributed to a significant academic study titled, "A Joint Convolutional-Recurrent Neural Network with an Attention Mechanism for Detecting Intracranial Hemorrhage on Noncontrast Head CT." This research investigates the efficacy of combining convolutional neural networks (CNN) with recurrent neural networks (RNN) and an attention mechanism in the detection and classification of intracranial hemorrhage (ICH) on a large-scale, multi-center dataset.

The study involved an extensive review of 55,179 head CT scans from 48,070 patients, which included 28,253 men, with an average age of 53.84 years. The development dataset was meticulously divided into training and validation sets, with the latter comprising 5,211 head CT scans, 991 of which were annotated as ICH-positive. Impressively, the model achieved a binary accuracy of 99.41%, sensitivity of 99.70%, and specificity of 98.91% on the validation set.

Following the development phase, the model was integrated into the PACS environment of an independent center for over six months, assessing its performance in a real clinical setting on a prospective independent sample. This phase involved 452 head CT scans, yielding an accuracy of 96.02% and an average prediction time of approximately 45 seconds.

This research not only highlights the potential of advanced neural network architectures in medical imaging but also sets a benchmark for future AI implementations in clinical diagnostics, proving that such technology can be seamlessly integrated into existing radiology workflows and provide rapid, accurate diagnostic decisions.

For details: https://www.nature.com/articles/s41598-022-05872-x