Deep Learning Algorithm for Traumatic Splenic Injury Detection and Sequential Localization

Background

Splenic injury is the most common solid visceral injury in blunt abdominal trauma, and high-resolution abdominal computed tomography (CT) can adequately detect the injury. However, these lethal injuries sometime have been overlooked in current practice. Deep learning algorithms have proven their capabilities in detecting abnormal findings in medical images. The aim of this study is to develop a three-dimensional, unsupervised deep learning algorithm for detecting splenic injury on abdominal CT using a sequential localization and classification approach.

Material and Methods

We retrospectively collected data from patients who underwent contrast-enhanced abdominal CT in the emergency department of Chang Gung Memorial Hospital, Linko, due to trauma and acute abdomen from Jul 2008 to Dec 2017. All patients were registered in the trauma and acute abdomen registries. Demographic information, including age, sex, disease diagnosis, trauma mechanism, Injury Severity Score, New Injury Severity Score , Abbreviated Injury Scale, and spleen injury grade, was collected. Scans showing splenic injury were identified as positive, and the remaining scans were defined as negative controls. We identified 300 venous phase scans with splenic injury and randomly selected 300 additional venous phase scans from the negative controls. CT scans with abdominal trauma injuries other than splenic injury were not excluded to reduce the selection bias. All data were split by age, sex, the presence of splenic injury, and injury severity score using stratified sampling into the developmental dataset and the test set at a ratio of 8:2. One-eighth of the developmental dataset was further reserved as the validation set during model construction.

Image preprocessing and labeling

The CT scan images were acquired in the original Digital Imaging and Communications in Medicine (DICOM) format. The images were then converted to the Neuroimaging Informatics Technology Initiative format, producing 3D voxel-based images. Our algorithm was then developed based on the venous axial slices, the most common imaging direction in abdominal trauma surveys. During the training process, image augmentation by translation, rotation, scaling, and elastic distortions was applied to increase model generalizability.

A trauma surgeon with 10 years of experience confirmed all the positive and negative scans. In all scans, the spleen with its surrounding background was covered with a manually drawn 3D bounding box.

Spleen localization

The localization model was designed based on 3D Faster RCNN with Resnet-101as the backbone structure and trained on the development dataset. We used cross-entropy, focal loss as the class loss, and L1 loss, distance intersection over union (DIOU) as box regression loss, and adopted intersection over union (IOU) and DIOU in non-maximum suppression (NMS) for training of the object detection algorithm.

Spleen injury identification and visualization

The cropped 3D images were used to develop the splenic injury classification model. We modified the block architecture to improve the interpretability of the reasoning process of the learned network. The output of the model was the probability of splenic injury.

Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, positive predictive value ,and negative predictive value.