Deep Learning Models Can Support Clinical Suspicion of Abusive Head Injury | Image credit: © Shisu_ka – © Shisu_ka – Stock.adobe.com.
Retinal hemorrhage (RH) is considered “strong evidence” of abusive head injury (AHT) in children, according to a recently published diagnostic study, but is missed by routine computed tomography (CT) It is said that there are many things. Jama network open, The investigators found that deep learning-based ocular analysis on pediatric head CT can predict the presence of RH. Deep learning models embedded in CT image analysis software provide support for adjusting clinical suspicion of AHT and determining which patients require urgent fundus examination.
CT is commonly used in the emergency department (ED) to rule out various intracranial abnormalities in infants and young children. According to the study authors, the image processing currently cannot identify RH unless it is “exceptionally” large. Identification of RH is an important part of AHT evaluation, which requires dilated fundoscopy, but this test is less widespread. The test may require sedation, which temporarily disables pupillary response. As a result, dilated ophthalmoscopy is “only performed in patients with the highest potential for abuse,” the authors write. AHT in infants and young children is associated with 40% severe disability and 25% mortality. Despite patient assessment, 25% to 31% of AHT in this population may have a wide spectrum of symptoms (overlapping with common childhood illnesses) and potentially misleading medical histories. % is missed.
It is recognized that deep learning has the potential to contribute to AHT imaging through predictive analytics, image analysis, and clinical decision support. The researchers hypothesized that deep-learning-based analysis of pediatric head CT spheres could predict the presence of RH because computer vision can “distinguish features not visible to human visual inspection.” I put it up.
To increase homogeneity of eye size and developmental stage, the study population consisted of 301 patients younger than 3 years of age. The median age was 4.6 (0.1-35.8) months, and 187 patients (62.1%) were male. Participants were diagnosed with AHT between 1 May 2007 and 31 March 2021 by Le Bonheur Children’s Hospital and its child abuse team. Diagnosis was based on physical examination, dilated fundus examination, imaging and clinical examination, medical history, and “other necessary examinations.” The presence or absence of RH was the outcome label for each sphere.
To evaluate the deep learning model, this study used axial slices from 218 segmented globes with RH and 384 segmented globes without RH. Two additional light gradient boosting machine models (GBMs) were also evaluated. One model used general brain findings in AHT and demographic features, and the other model “combined the same demographic features and brain findings with deep learning model risk prediction.” . Sensitivity (recall), accuracy, precision, specificity, F1 score, and area under the curve (AUC) were evaluated to predict the presence or absence of her RH in the eye. The regions of the Earth that influenced the predictions of the deep learning model were visualized in a saliency map, and the contributions of “demographic features and standard his CT features were evaluated by Shapley’s additive description of his ”.
In this study, 120 (39.3%) of patients had RH on fundus examination. Below are the performance results of the deep learning model: Sensitivity, 79.6%. Specificity, 79.2%. Accuracy, 68.6%. Negative predictive value, 87.1%. Accuracy, 79.3%. F1 score, 73.7%. AUC, 0.83 (95% CI, 0.79-0.93). For the common mild GBM model, the AUC was 0.80 (95% CI, 0.69 to 0.91). The AUC for the combined optical GBM model was 0.86 (95% CI, 0.79–0.93). The specificity of combining the lightweight GBM model and the deep learning model was higher compared to the lightweight GBM model. Sensitivity was similar for all models, according to the authors.
In this diagnostic study, the authors concluded that RH information is present in head CT and accessible through deep learning image analysis. Identifying RH on head CT allows clinicians practicing in a subspecialty setting to “use a routine diagnostic tool that is objective and not susceptible to common clinical biases. We can be very confident that we will advance AHT research and reduce the number of missed cases.” ”
reference:
Gunturkun F, Bakir-Batu B, Siddiqui A, et al. Development of a deep learning model for retinal hemorrhage detection in infant head computed tomography.JAMA net opened.2023;6(6):e2319420. doi:10.1001/jamanetworkopen.2023.19420
