In this study, we developed an algorithm that can detect injuries in five abdominal organs: liver, spleen, kidney, intestine, and extravasation. Our results demonstrate that the algorithm can accurately diagnose parenchymal organ injuries. The algorithm can localize the abdominal organs and then detect the injuries in each organ simultaneously, which can assist clinicians in efficient screening and triage, facilitate the treatment of trauma patients, and avoid the waste of medical resources.

Furthermore, the algorithm performed best in identifying kidney injury on abdominal CT scans, with an ACC of 0.932 (PPV: 0.888; NPV: 0.943; Sensitivity: 0.887; Specificity: 0.944). It also showed good performance in diagnosing liver and spleen injuries, with an ACC of 0.873 (PPV: 0.789; NPV: 0.895; Sensitivity: 0.789; Specificity: 0.895) and 0.771 (PPV: 0.63; NPV: 0.814; Sensitivity: 0.626; Specificity: 0.816), respectively.

In clinical practice, radiologists’ diagnostic focus and efficiency can be impacted by various factors, including fatigue and time pressure. Studies have shown that the error rates of radiologists in abdominal CT diagnosis fluctuate throughout the day and week. Specifically, during the workweek, error rates are highest later in the morning and at the end of the workday, with Mondays showing higher rates compared to other days [17]. Additionally, there are notable variations in expertise and experience among radiologists of different ages and qualifications. Studies have found that less experienced radiologists may have error rates as high as 32% in diagnosing abdominal solid organ CT images under busy conditions [18,19,20]. In contrast, diagnostic models based on deep learning exhibit robust stability and rapid speed, unaffected by subjective or objective factors, which can operate continuously for 24 h a day. During our study, we uploaded 3,147 patients’ CT images for the model learning, which cost about 5 h until the diagnosis completion. For individual patient, it takes seconds or minutes to finish the analysis, the time depends on the difference and quality of each CT image.

Previous studies have applied deep learning algorithms to diagnose specific abdominal injuries, such as kidney segmentation [21], splenic laceration [22], liver laceration [23, 24], and abdominal hemorrhage [25]. However, none of these studies have attempted to detect multiple organ injuries in trauma patients. Therefore, we developed a deep learning algorithm that can detect injuries in five different abdominal organs at once. We used a 2D semantic segmentation model to extract the organs from the CT images, and then a 2.5D classification model to predict the injury probability of each organ. This approach improved the speed and accuracy of the algorithm.

We developed an algorithm that requires large amounts of accurately labeled data to achieve high performance, to facilitate the labeling process for the clinicians and enhance the results of the automatic detection algorithm. The dataset for CT examinations in this study included both conventional and enhanced CT scans. We conducted a normalization on the data to avoid the effect of the contrast agent usage in patient before uploaded to the model. A 2.5D classification model had been introduced instead of a 3D model, which could recognize the small data sets, for reduction parameters numbers and prevention the overfitting while preserving the model performance. The LSTM was included as well, for the components could performed spatial analysis on the input serial CT images and capture the spatial relationships among the images.

To improve the diversity and quality of the data set and address the data imbalance issue, we applied various data augmentation techniques that mimic different scenarios that can affect the quality of CT images and help the model cope with reality variations. With the basic geometric transformations, such as horizontal and vertical flipping, which helped to distinguish the same anatomical structures in different orientation caused by scanning angles.

The Geometric transformation technology, blur technology and random Gaussian noise was applied for data enhancement, which can effectively enrich the training data set, resist the occurrence of over-fitting, and enable the model to make correct fitting when facing new images, instead of blindly limiting it to some known images.

Over all, deep learning has been widely applied to clinical data analysis, especially in image processing [26]. It has advanced the field of medical imaging by enabling the identification, classification, and quantification of patterns in various modalities [27], and quantitative assessment of blunt liver trauma in children [28], which can provide clinicians with accurate and fast diagnostic assistant [29]. Our deep learning model provided a high-quality image analysis result that helps clinicians perform quick screening and triage, identify patients with abdominal trauma, to improve medical efficiency and save medical resources when in natural disasters or mass accidents. Moreover, the model has the potential to be applied to the CT diagnosis of other diseases.