Deep learning models have the potential to assist clinicians in determining suspicion for abusive head trauma (AHT) by screening pediatric head computed tomography (CT) images for retinal hemorrhage, according to a recent study.
Retinal hemorrhage (RH) is considered “strong evidence” for abusive head trauma (AHT) in children, but often missed in routine computed tomography (CT), according to a recently published diagnostic study in Jama Network Open, which found deep learning-based analysis of globes on pediatric head CTs can predict RH presence. Incorporated into CT image analysis software, a deep learning model can calibrate clinical suspicion for AHT, providing support to determine which patients urgently need fundoscopic examinations.
To exclude a range of intracranial abnormalities in infants and young children, CT is commonly used in emergency departments (ED). With this imaging, RH cannot currently be identified unless they are “exceptionally” large, according to the study authors. Identifying RH is an essential part of an AHT assessment, which requires a dilated fundoscopic examination, a subspecialty that isn’t widely available. The examination could require sedation and temporarily nullifies pupillary response. As a result, dilated fundoscopic examinations are “reserved for those patients with the highest likelihood of abuse,” the authors wrote. AHT in infants and young children is associated with 40% severe disability and 25% mortality. Since patients with AHT can present with wide-ranging symptoms (that overlap with common pediatric illnesses) and a potentially misleading history, 25% to 31% of AHTs in this population are missed, despite patient evaluation.
It is recognized that there is potential for deep learning to contribute to diagnostic imaging for AHT through predictive analytics, image analysis, and clinical decision support. Investigators hypothesized that deep learning-based analysis of pediatrics head CT globes could predict presence and absence of RH, as computer vision can “discern features that are otherwise inapparent to human visual examination.”
The study population was made up of 301 patients younger than 3 years to increase uniformity in globe size and stage of development. The median age was 4.6 (0.1-35.8) months, and 187 patients were male (62.1%). Participants were diagnosed with AHT from May 1, 2007, to March 31, 2021, by Le Bonheur Children’s Hospital and its child abuse team. Diagnoses were made based on physical examination, dilated fundoscopic examinations, imaging and laboratory studies, history, and “other necessary investigations.” The presence or absence of RHs was the outcome label for each globe.
To assess the deep learning model, the study used axial slices from 218 segmented globes with RH and 384 globes without RH. Also assessed were two additional light gradient boosting machine models (GBM). One model used common brain findings in AHT and demographic characteristics, while the other “combined the deep learning model’s risk prediction plus the same demographic characteristics and brain findings.” To predict the presence or absence of RH in the globes, sensitivity (recall), precision, accuracy, specificity, F1 score, and area under the curve (AUC) were assessed. Globe regions that influenced predictions of the deep learning model were visualized in saliency maps, and “contributions of demographic and standard CT features were assessed by Shapley additive explanation.”
In the study, 120 (39.3%) of patients had RH on fundoscopic examinations. The following are results of the deep learning model’s performance: sensitivity, 79.6%; specificity, 79.2%; precision, 68.6%; negative predictive value, 87.1%; accuracy, 79.3%; F1 score, 73.7%; and AUC, 0.83 (95% CI, 0.79-0.93). For the general light GBM model, AUCs were 0.80 (95% CI, 0.69-0.91). AUCs for the combined light GBM model were 0.86 (95% CI, 0.79-0.93). Specificities for the combined light GBM and deep learning models were higher compared to the light GBM model. In all models, sensitivities were similar, according to the authors.
In the diagnostic study, authors concluded RH information is present in head CTs, which can be accessed through deep learning image analysis. By discriminating RH on head CTs, clinicians practicing in subspecialty-limited environments could have greater confidence in “moving an AHT investigation forward and decrease the number of missed cases, all by using a routine diagnostic modality that is objective and less susceptible to common clinical bias.”
Gunturkun F, Bakir-Batu B, Siddiqui A, et al. Development of a Deep Learning Model for Retinal Hemorrhage Detection on Head Computed Tomography in Young Children. JAMA Netw Open. 2023;6(6):e2319420. doi:10.1001/jamanetworkopen.2023.19420