A single artificial intelligence model is asked to do two important jobs on a stroke CT angiogram: find a blocked large vessel and score the brain’s collateral blood supply. In an external test, it correctly classified 339 of 345 scans. According to the World Health Organization, 15 million people suffer from a stroke attack every year, resulting in 6 million deaths and 5 million disabilities.

Ischemic stroke with large vessel occlusion, or LVO, is the most severe type of ischemic stroke, and approximately one-third of ischemic strokes are caused by LVO. Mechanical thrombectomy is the standard treatment for acute ischemic stroke with LVO, because several large-scale studies in 2015 demonstrated its benefits in restoring blood flow in eligible patients.

Early diagnosis of LVO is crucial for the success of mechanical thrombectomy, because the chances of achieving functional independence decrease with delays. Computed tomography angiography, or CTA, has become the standard for acute stroke imaging in assessing LVO, and it also provides information about collateral status useful for decision-making in mechanical thrombectomy.

The severe consequences of delayed LVO diagnosis and inter-observer variations among radiologists in assessing collateral status necessitate automated diagnosis. The study approached detection of LVO and collateral scoring on CTA as a multi-task three-dimensional object detection problem, using a state-of-the-art self-configuring nnDetection model on a large-scale, multi-center, multi-vendor data set.

The model was used to identify LVO and assess collateral status on single-phase CTA, then its performance was assessed on external data. The dataset was divided into a development sample for training and validating the deep learning models, and an external test sample for testing their unbiased performance in identifying LVO and assigning collateral scores.

CTA scans from five out of six centers formed the development set, while the remaining center’s CTA scans formed the test set. The development scans came from Siemens, GE Healthcare, Philips, and Canon scanners, while the test set came from Siemens and Canon scanners.

The maximum slice thickness was one point five millimeters in the development set and one millimeter in the test set. Figure one shows the study flow from computed tomography angiography scans collected at Centers one through five into a development set, after excluding severe motion artifacts, absent contrast opacification, and patients younger than eighteen.

The development set contained two thousand four hundred twenty-five patients, including four hundred thirty-five with large vessel occlusion. The authors trained and validated their nnDetection model on this set, then assessed its performance on external test data from Center six, providing an evaluation across a separate clinical site.

The study used nnDetection, a state-of-the-art three-dimensional self-configuring medical object detection deep learning method. nnDetection selects the best architecture depending on the dataset by using a data fingerprint, rule-based parameters, fixed parameters, and empirical parameter optimization.

The nnDetection model was built on top of Retina U-net, using a similar network topology. Figure two shows nnDetection’s feature-pyramid architecture, combining a bottom-up pathway that builds increasingly rich, lower-resolution features with a top-down pathway that restores spatial detail through skip connections.

The orange feature maps, labeled P5 through P2, feed classification and bounding-box heads for detecting the middle cerebral artery zone and large-vessel occlusion. For patients with occlusion, an auxiliary branch identifies the brain hemisphere and supports collateral assessment.

A total of 2,425 patients were enrolled in the development sample; 1,746 were men, or 72 percent, and the mean age was 55.74 years, with a range from 30 to 90. There were 435 LVO-positive patients, representing 17.79 percent of the development sample.

Radiologists’ consensus assigned good collateral status to 200 patients and poor collateral status to 235 patients in the development set. On the external test set, the nnDetection model achieved 98.26 percent diagnostic accuracy for identifying LVO, correctly classifying 339 of 345 CTA scans.

Its F1-score and sensitivity were both 96.30 percent, while specificity and negative predictive value were both 98.86 percent, and positive predictive value was 96.30 percent. Of the 345 scans, 165, or 47.82 percent, were arterial phase, and 180, or 52.18 percent, were venous phase.

For arterial-phase scans, the model’s F1-score was 94.24 percent, accuracy was 96.97 percent, sensitivity was 93.18 percent, specificity was 98.35 percent, negative predictive value was 97.54 percent, and positive predictive value was 95.35 percent. For venous-phase scans, the corresponding values were 98.59 percent, 99.44 percent, 100 percent, 99.3 percent, 100 percent, and 97.37 percent.

Figure three shows three examples where the deep learning model made correct predictions on CTA images. Dashed boxes mark the model’s predictions, while continuous boxes show the ground truth; red identifies the pathological side and green the normal side.

The examples include left middle cerebral artery M1 segment occlusion, poor collateral status with right-sided occlusion, and good collateral status with left-sided occlusion, illustrating how object detection provides an interpretable visual output. Figure four shows three representative failure cases for the object-detection model.

In panel a, a large intraparenchymal hematoma displaces the right middle cerebral artery, likely obscuring normal contrast filling and leading to a false LVO prediction; panel b shows a missed left proximal M2 occlusion. Panel c shows an occluded left M1 segment that the model labeled as having good collaterals, although experts judged them poor.

These examples make the model’s clinically important error modes visible. Most false-negative cases were M2 occlusions, which are inherently difficult to detect, even for less-experienced radiologists. False-positive cases primarily arose because of accompanying pathologies, such as intracerebral hematomas.

Suggested ways to address these issues include adding more distal occlusion cases to the training dataset, developing a specialized model for challenging cases, and increasing false-positive inducers in the training data. Safeguards could also identify hematomas and halt subsequent LVO detection or alert users to potential risks, which could help improve system accuracy.

In conclusion, a self-configuring three-dimensional nnDetection model can accurately identify LVO on single-phase CTA scans, providing the occlusion location with understandable bounding boxes. Using a multi-tasking approach, the same model can accurately provide semi-quantitative collateral scores, offering a one-stop-shop approach for automated stroke diagnostics on CTA in patients with LVO.

The self-configuring three-dimensional nnDetection model accurately identified large vessel occlusion and provided semi-quantitative collateral scores on single-phase CTA, suggesting a one-stop automated approach for stroke diagnostics.