Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study.

BACKGROUND: Non-contrast head CT scan is the current standard for initial imaging of patients with head trauma or stroke symptoms. We aimed to develop and validate a set of deep learning algorithms for automated detection of the following key findings from these scans: intracranial haemorrhage and its types (ie, intraparenchymal, intraventricular, subdural, extradural, and subarachnoid); calvarial fractures; midline shift; and mass effect. METHODS: We retrospectively collected a dataset containing 313 318 head CT scans together with their clinical reports from around 20 centres in India between Jan 1, 2011, and June 1, 2017. A randomly selected part of this dataset (Qure25k dataset) was used for validation and the rest was used to develop algorithms. An additional validation dataset (CQ500 dataset) was collected in two batches from centres that were different from those used for the development and Qure25k datasets. We excluded postoperative scans and scans of patients younger than 7 years. The original clinical radiology report and consensus of three independent radiologists were considered as gold standard for the Qure25k and CQ500 datasets, respectively. Areas under the receiver operating characteristic curves (AUCs) were primarily used to assess the algorithms. FINDINGS: The Qure25k dataset contained 21 095 scans (mean age 43 years; 9030 [43%] female patients), and the CQ500 dataset consisted of 214 scans in the first batch (mean age 43 years; 94 [44%] female patients) and 277 scans in the second batch (mean age 52 years; 84 [30%] female patients). On the Qure25k dataset, the algorithms achieved an AUC of 0·92 (95% CI 0·91-0·93) for detecting intracranial haemorrhage (0·90 [0·89-0·91] for intraparenchymal, 0·96 [0·94-0·97] for intraventricular, 0·92 [0·90-0·93] for subdural, 0·93 [0·91-0·95] for extradural, and 0·90 [0·89-0·92] for subarachnoid). On the CQ500 dataset, AUC was 0·94 (0·92-0·97) for intracranial haemorrhage (0·95 [0·93-0·98], 0·93 [0·87-1·00], 0·95 [0·91-0·99], 0·97 [0·91-1·00], and 0·96 [0·92-0·99], respectively). AUCs on the Qure25k dataset were 0·92 (0·91-0·94) for calvarial fractures, 0·93 (0·91-0·94) for midline shift, and 0·86 (0·85-0·87) for mass effect, while AUCs on the CQ500 dataset were 0·96 (0·92-1·00), 0·97 (0·94-1·00), and 0·92 (0·89-0·95), respectively. INTERPRETATION: Our results show that deep learning algorithms can accurately identify head CT scan abnormalities requiring urgent attention, opening up the possibility to use these algorithms to automate the triage process. FUNDING: Qure.ai.

SkullBreak / SkullFix - Dataset for automatic cranial implant design and a benchmark for volumetric shape learning tasks.

The article introduces two complementary datasets intended for the development of data-driven solutions for cranial implant design, which remains to be a time-consuming and laborious task in current clinical routine of cranioplasty. The two datasets, referred to as the SkullBreak and SkullFix in this article, are both adapted from a public head CT collection CQ500 (http://headctstudy.qure.ai/dataset) with CC BY-NC-SA 4.0 license. The SkullBreak contains 114 and 20 complete skulls, each accompanied by five defective skulls and the corresponding cranial implants, for training and evaluation respectively. The SkullFix contains 100 triplets (complete skull, defective skull and the implant) for training and 110 triplets for evaluation. The SkullFix dataset was first used in the MICCAI 2020 AutoImplant Challenge (https://autoimplant.grand-challenge.org/) and the ground truth, i.e., the complete skulls and the implants in the evaluation set are held private by the organizers. The two datasets are not overlapping and differ regarding data selection and synthetic defect creation and each serves as a complement to the other. Besides cranial implant design, the datasets can be used for the evaluation of volumetric shape learning algorithms, such as volumetric shape completion. This article gives a description of the two datasets in detail.

Automated Lateral Ventricular and Cranial Vault Volume Measurements in 13,851 Patients Using Deep Learning Algorithms.

BACKGROUND: No large dataset-derived standard has been established for normal or pathologic human cerebral ventricular and cranial vault volumes. Automated volumetric measurements could be used to assist in diagnosis and follow-up of hydrocephalus or craniofacial syndromes. In this work, we use deep learning algorithms to measure ventricular and cranial vault volumes in a large dataset of head computed tomography (CT) scans. METHODS: A cross-sectional dataset comprising 13,851 CT scans was used to deploy U-Net deep learning networks to segment and quantify lateral cerebral ventricular and cranial vault volumes in relation to age and sex. The models were validated against manual segmentations. Corresponding radiologic reports were annotated using a rule-based natural language processing framework to identify normal scans, cerebral atrophy, or hydrocephalus. RESULTS: U-Net models had high fidelity to manual segmentations for lateral ventricular and cranial vault volume measurements (Dice index, 0.878 and 0.983, respectively). The natural language processing identified 6239 (44.7%) normal radiologic reports, 1827 (13.1%) with cerebral atrophy, and 1185 (8.5%) with hydrocephalus. Age-based and sex-based reference tables with medians, 25th and 75th percentiles for scans classified as normal, atrophy, and hydrocephalus were constructed. The median lateral ventricular volume in normal scans was significantly smaller compared with hydrocephalus (15.7 vs. 82.0 mL; P < 0.001). CONCLUSIONS: This is the first study to measure lateral ventricular and cranial vault volumes in a large dataset, made possible with artificial intelligence. We provide a robust method to establish normal values for these volumes and a tool to report these on CT scans when evaluating for hydrocephalus.