Detection of Intracranial Hemorrhage for Trauma Patients

TL;DR

This paper tackles the challenge of rapidly detecting intracranial hemorrhage in whole-body trauma CT scans by framing it as 3D bounding-box detection rather than full segmentation. It introduces a 3D Retina-Net with anisotropic multiscale features and a novel Voxel-Complete IoU (VC-IoU) loss that enforces 3D aspect-ratio consistency across axes, improving localization of bleedings with diverse shapes and sizes. Evaluated on INSTANCE2022 and a private external cohort, the method achieves about a 5% relative gain in Average Recall over strong baselines and demonstrates robustness to dataset shift, while also analyzing annotation strategies and the impact of noisy labels. These results suggest practical utility for aiding radiologists in reducing missed hemorrhages and inform cost-effective annotation choices for 3D medical object detection in data-limited clinical settings.

Abstract

Whole-body CT is used for multi-trauma patients in the search of any and all injuries. Since an initial assessment needs to be rapid and the search for lesions is done for the whole body, very little time can be allocated for the inspection of a specific anatomy. In particular, intracranial hemorrhages are still missed, especially by clinical students. In this work, we present a Deep Learning approach for highlighting such lesions to improve the diagnostic accuracy. While most works on intracranial hemorrhages perform segmentation, detection only requires bounding boxes for the localization of the bleeding. In this paper, we propose a novel Voxel-Complete IoU (VC-IoU) loss that encourages the network to learn the 3D aspect ratios of bounding boxes and leads to more precise detections. We extensively experiment on brain bleeding detection using a publicly available dataset, and validate it on a private cohort, where we achieve 0.877 AR30, 0.728 AP30, and 0.653 AR30, 0.514 AP30 respectively. These results constitute a relative +5% improvement in Average Recall for both datasets compared to other loss functions. Finally, as there is little data currently publicly available for 3D object detection and as annotation resources are limited in the clinical setting, we evaluate the cost of different annotation methods, as well as the impact of imprecise bounding boxes in the training data on the detection performance.

Figures (8)

• Figure 1: Insights into the challenges of brain hemorrhage detection with examples of hard cases. A) The patient suffers both from a bleeding and a thrombosis further below. While the thrombosis has a similar appearance to the bleeding, the blood is still in a blood vessel and is not considered a bleeding. B and C) Both images show hyperdensities near the ventricle system. While the spot in image B is a bleeding, the two spots in C are simple calcifications of the brain tissue.
• Figure 2: Our method for bleeding detection. Anchor sizes are given for level P2.
• Figure 3: In 2D, the aspect ratio $\phi$ gives a scale-independent representation of a given bounding box. However, no such representation exists for 3D bounding boxes. We can instead consider the three aspect ratios $\{\phi_{ax}, \phi_{sag}, \phi_{cor}\}$ in the three natural planes of the 3D space (axial, sagittal and coronal). In particular, bleeding have a wide variety of shapes even within a single individual.
• Figure 4: Distribution shift of the number of bleeding per image (left) and volume per bleeding (right) between the INSTANCE2022 dataset and our private cohort.
• Figure 5: Visualization of head CTs, axial (top) and sagittal (bottom). INSTANCE2022 data (left). Downscaling (middle) of the original private cohort data (right).