Validation of a machine learning software tool for automated large vessel occlusion detection in patients with suspected acute stroke.

PURPOSE: CT angiography (CTA) is the imaging standard for large vessel occlusion (LVO) detection in patients with acute ischemic stroke. StrokeSENS LVO is an automated tool that utilizes a machine learning algorithm to identify anterior large vessel occlusions (LVO) on CTA. The aim of this study was to test the algorithm's performance in LVO detection in an independent dataset. METHODS: A total of 400 studies (217 LVO, 183 other/no occlusion) read by expert consensus were used for retrospective analysis. The LVO was defined as intracranial internal carotid artery (ICA) occlusion and M1 middle cerebral artery (MCA) occlusion. Software performance in detecting anterior LVO was evaluated using receiver operator characteristics (ROC) analysis, reporting area under the curve (AUC), sensitivity, and specificity. Subgroup analyses were performed to evaluate if performance in detecting LVO differed by subgroups, namely M1 MCA and ICA occlusion sites, and in data stratified by patient age, sex, and CTA acquisition characteristics (slice thickness, kilovoltage tube peak, and scanner manufacturer). RESULTS: AUC, sensitivity, and specificity overall were as follows: 0.939, 0.894, and 0.874, respectively, in the full cohort; 0.927, 0.857, and 0.874, respectively, in the ICA occlusion cohort; 0.945, 0.914, and 0.874, respectively, in the M1 MCA occlusion cohort. Performance did not differ significantly by patient age, sex, or CTA acquisition characteristics. CONCLUSION: The StrokeSENS LVO machine learning algorithm detects anterior LVO with high accuracy from a range of scans in a large dataset.

Evaluating nnU-Net for early ischemic change segmentation on non-contrast computed tomography in patients with Acute Ischemic Stroke.

Identifying the presence and extent of early ischemic changes (EIC) on Non-Contrast Computed Tomography (NCCT) is key to diagnosing and making time-sensitive treatment decisions in patients that present with Acute Ischemic Stroke (AIS). Segmenting EIC on NCCT is however a challenging task. In this study, we investigated a 3D CNN based on nnU-Net, a self-adapting CNN technique that has become the state-of-the-art in medical image segmentation, for segmenting EIC in NCCT of AIS patients. We trained and tested this model on a sizeable and heterogenous dataset of 534 patients, split into 438 for training and validation and 96 for testing. On this test set, we additionally assessed the inter-rater performance by comparing the proposed approach against two reference segmentation annotations by expert neuroradiologist readers, using this as the benchmark against which to compare our model. In terms of spatial agreement, we report median Dice Similarity Coefficients (DSCs) of 39.8% for the model vs. Reader-1, 39.4% for the model vs. Reader-2, and 55.6% for Reader-2 vs. Reader-1. In terms of lesion volume agreement, we report Intraclass Correlation Coefficients (ICCs) of 83.4% for model vs. Reader-1, 80.4% for model vs. Reader-2, and 94.8% for Reader-2 vs. Reader-1. Based on these results, we conclude that our model performs well relative to expert human performance and therefore may be useful as a decision-aid for clinicians.

Automatically Delineating Key Anatomy in 3-D Ultrasound Volumes for Hip Dysplasia Screening.

Developmental dysplasia of the hip (DDH) metrics based on 3-D ultrasound have proven more reliable than those based on 2-D images, but to date have been based mainly on hand-engineered features. Here, we test the performance of 3-D convolutional neural networks for automatically segmenting and delineating the key anatomical structures used to define DDH metrics: the pelvis bone surface and the femoral head. Our models are trained and tested on a data set of 136 volumes from 34 participants. For the pelvis, a 3D-U-Net achieves a Dice score of 85%, outperforming the confidence-weighted structured phase symmetry algorithm (Dice score = 19%). For the femoral head, the 3D-U-Net had centre and radius errors of 1.42 and 0.46 mm, respectively, outperforming the random forest classifier (3.90 and 2.01 mm). The improved segmentation may improve DDH measurement accuracy and reliability, which could reduce misdiagnosis.