Deep learning-based renal-trauma classification: comparison of two-dimensional convolutional neural network and vision-transformer models on arterial- and venous-phase computed-tomography images.

Authors

Abstract

Rapid and accurate detection of renal trauma is essential in emergency medicine. This study aimed to automatically classify renal trauma using 2D convolutional neural network (CNN) and vision transformer (ViT) models on arterial- and venous-phase computed tomography (CT) images. We analyzed CT images from 94 patients, including 47 without renal trauma and 47 with renal trauma (30,951 and 15,086 images, respectively). Arterial- and venous-phase CT images were reconstructed into axial, coronal, and sagittal images, with the renal regions cropped to construct the image datasets. Five pretrained CNN-based and five pretrained ViT-based 2D models were fine-tuned to classify renal trauma, and their performance was evaluated at the slice and patient levels using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. At the slice level, the sagittal Inception-V3 model achieved the highest AUC values of 0.986 for venous-phase and 0.975 for arterial-phase images. At the patient level, the coronal Inception-V3 model achieved the highest AUC values of 0.991 for venous-phase and 0.945 for arterial-phase images. CNN-based models outperformed ViT-based models at both levels, with slice-level mean AUC values of 0.971 and 0.940 for venous-phase and 0.942 and 0.914 for arterial-phase images and patient-level values of 0.947 and 0.928 for venous-phase and 0.914 and 0.886 for arterial-phase images, respectively. The 2D models demonstrated high diagnostic performance for renal trauma, even when using arterial-phase images alone, supporting their potential for rapid decision-making in emergency settings.

Tags

Topics