Evaluation of techniques to improve a deep learning algorithm for the automatic detection of intracranial haemorrhage on CT head imaging

Abstract

Background: Deep learning (DL) algorithms are playing an increasing role in automatic medical image analysis. Purpose: To evaluate the performance of a DL model for the automatic detection of intracranial haemorrhage and its subtypes on non-contrast CT (NCCT) head studies and to compare the effects of various preprocessing and model design implementations. Methods: The DL algorithm was trained and externally validated on open-source, multi-centre retrospective data containing radiologist-annotated NCCT head studies. The training dataset was sourced from four research institutions across Canada, the USA and Brazil. The test dataset was sourced from a research centre in India. A convolutional neural network (CNN) was used, with its performance compared against similar models with additional implementations: (1) a recurrent neural network (RNN) attached to the CNN, (2) preprocessed CT image-windowed inputs and (3) preprocessed CT image-concatenated inputs. The area under the receiver operating characteristic curve (AUC-ROC) and microaveraged precision (mAP) score were used to evaluate and compare model performances. Results: The training and test datasets contained 21,744 and 491 NCCT head studies, respectively, with 8,882 (40.8%) and 205 (41.8%) positive for intracranial haemorrhage. Implementation of preprocessing techniques and the CNN-RNN framework increased mAP from 0.77 to 0.93 and increased AUC-ROC [95% confidence intervals] from 0.854 [0.816–0.889] to 0.966 [0.951–0.980] (p-value = 3.91 × 10−12). Conclusions: The deep learning model accurately detected intracranial haemorrhage and improved in performance following specific implementation techniques, demonstrating clinical potential as a decision support tool and an automated system to improve radiologist workflow efficiency. Key points: • The deep learning model detected intracranial haemorrhages on computed tomography with high accuracy. • Image preprocessing, such as windowing, plays a large role in improving deep learning model performance. • Implementations which enable an analysis of interslice dependencies can improve deep learning model performance. • Visual saliency maps can facilitate explainable artificial intelligence systems. • Deep learning within a triage system may expedite earlier intracranial haemorrhage detection.