CT Scans As Video: Efficient Intracranial Hemorrhage Detection Using Multi-Object Tracking

TL;DR

The study reframes CT volumes as video streams to bridge 2D detector efficiency and 3D context for intracranial hemorrhage detection on resource-constrained devices. By integrating a YOLO-based slice detector with ByteTrack-inspired slice-to-slice tracking, a bi-directional processing scheme, and a lightweight hybrid inference plus a spatiotemporal consistency filter, the approach achieves higher Precision (up to $0.779$) while maintaining Recall relative to a baseline 2D detector. Key contributions include the novel video-viewpoint CT paradigm, a Hybrid tracking strategy that overcomes initialization lag, and practical evidence from the Hemorica dataset that this method enables scalable, real-time triage for edge-deployed stroke care. The findings suggest that explicit 3D reasoning can be approximated with efficient 2D tools, enabling deployment in mobile clinics and IoT-enabled scanners, with future work focusing on domain adaptation and improved occlusion handling via appearance-based re-identification features.

Abstract

Automated analysis of volumetric medical imaging on edge devices is severely constrained by the high memory and computational demands of 3D Convolutional Neural Networks (CNNs). This paper develops a lightweight computer vision framework that reconciles the efficiency of 2D detection with the necessity of 3D context by reformulating volumetric Computer Tomography (CT) data as sequential video streams. This video-viewpoint paradigm is applied to the time-sensitive task of Intracranial Hemorrhage (ICH) detection using the Hemorica dataset. To ensure operational efficiency, we benchmarked multiple generations of the YOLO architecture (v8, v10, v11 and v12) in their Nano configurations, selecting the version with the highest mAP@50 to serve as the slice-level backbone. A ByteTrack algorithm is then introduced to enforce anatomical consistency across the $z$-axis. To address the initialization lag inherent in video trackers, a hybrid inference strategy and a spatiotemporal consistency filter are proposed to distinguish true pathology from transient prediction noise. Experimental results on independent test data demonstrate that the proposed framework serves as a rigorous temporal validator, increasing detection Precision from 0.703 to 0.779 compared to the baseline 2D detector, while maintaining high sensitivity. By approximating 3D contextual reasoning at a fraction of the computational cost, this method provides a scalable solution for real-time patient prioritization in resource-constrained environments, such as mobile stroke units and IoT-enabled remote clinics.

Figures (3)

• Figure 1: Overview of the proposed video-based detection pipeline.
• Figure 2: Training Dynamics. Evolution of Box Loss (Red) and mAP@50 (Blue) over 50 epochs. The validation loss closely tracks the training loss, confirming stable convergence without overfitting.
• Figure 3: Qualitative comparison across methods. We visualize consecutive slices (Columns) to demonstrate temporal consistency. Row 1 (Green): Ground Truth annotations. Row 2 (Yellow): Baseline YOLOv11n detections showing slice-level inconsistencies. Row 3 (Blue): ByteTrack results, showing track initialization lag. Row 4 (Purple): Proposed Hybrid method. Row 5 (Teal): Spatiotemporal Filter results. Note how the proposed methods (Rows 4-5) recover the missed detections in the middle columns compared to the baseline.