TLDR: OccluNet is a new deep learning model designed to automatically detect vascular occlusions in Digital Subtraction Angiography (DSA) images, crucial during stroke treatment. Unlike previous methods that often ignored the dynamic, temporal aspect of DSA, OccluNet integrates a YOLOX object detector with transformer-based temporal attention. This allows it to analyze the flow of contrast over time, significantly improving the accuracy of occlusion detection compared to traditional methods. The model achieved high precision and recall, demonstrating the critical role of spatio-temporal information in identifying blockages in stroke patients.
Acute Ischemic Stroke (AIS) is a life-threatening medical emergency caused by a sudden interruption of blood flow to the brain due to a blocked blood vessel. A common treatment for AIS is Endovascular Thrombectomy (EVT), a minimally invasive procedure to remove the blockage and restore blood flow. During EVT, doctors use Digital Subtraction Angiography (DSA) images to visualize the occlusion and guide the procedure.
However, interpreting DSA sequences can be challenging due to the complex anatomy of blood vessels and the time-sensitive nature of stroke treatment. Traditional methods for detecting occlusions in DSA often treat each image frame independently or combine them into a single view, which overlooks the crucial temporal information—how the contrast agent flows through the vessels over time.
Introducing OccluNet: A Spatio-Temporal Approach
To address these challenges, researchers have developed OccluNet, a novel deep learning model designed to automate the detection of vascular occlusions in DSA sequences. OccluNet stands out by integrating a powerful object detector called YOLOX with transformer-based temporal attention mechanisms. This allows the model to not only identify occlusions in individual frames but also to understand the dynamic flow patterns across a sequence of images, much like how a neurointerventionalist would scroll through images to spot blockages.
The model explores two main ways to incorporate temporal information: ‘pure temporal attention,’ which processes features across frames at each spatial location, and ‘divided space-time attention,’ which alternates between spatial and temporal analysis. Both approaches aim to capture the ‘where’ and ‘when’ of an occlusion.
How OccluNet Works
OccluNet takes a DSA sequence as input, considering the current frame along with its immediate neighbors. It first extracts spatial features from these frames using a YOLOX-based backbone. These features then pass through a temporal module, where the attention mechanisms come into play, enhancing the features with context from neighboring frames. Finally, a YOLOX head uses these enriched spatio-temporal features to pinpoint and classify occlusions with bounding boxes.
To ensure consistency over time, OccluNet employs a trajectory-based optimization. This links detections across multiple frames, rewarding predictions that are persistent over time and suppressing inconsistent false positives. This mimics the clinical observation that a true occlusion will appear consistently across a sequence of images.
Performance and Impact
OccluNet was evaluated using DSA images from the MR CLEAN Registry, a large dataset of acute ischemic stroke patients. The model significantly outperformed baseline models that did not incorporate temporal information. For instance, OccluNet1 (temporal attention variant) achieved an overall precision of 89.02% and a recall of 74.87%, demonstrating its strong capability in detecting occlusions. Both temporal attention variants of OccluNet showed similar high performance, indicating that the integration of temporal information is key, regardless of the specific attention mechanism used.
These findings align with clinical practice, where the dynamic nature of DSA is crucial for accurate diagnosis. Despite being trained with fewer epochs due to computational constraints compared to the baseline, OccluNet’s superior performance highlights the critical importance of leveraging temporal dynamics in DSA processing.
Also Read:
- AI Tool Improves Image Quality for Acute Ischemic Stroke Treatment
- AI System Enhances Real-time Brain Vessel Segmentation for Stroke Risk Assessment
Future Directions
While OccluNet represents a significant advancement, future work could explore more complex backbones for even greater power, integrate anatomical priors like vessel segmentation masks to reduce false positives, or incorporate multimodal inputs such as CTA scans. Expanding the dataset to include more diverse and rare occlusion types, especially distal occlusions which are harder to detect, will also be vital for improving the model’s generalization and clinical applicability.
OccluNet lays a strong foundation for new tools that could significantly improve clinical workflows during endovascular stroke interventions, potentially assisting radiologists in making faster and more accurate decisions. For more detailed information, you can refer to the full research paper here.
