EP540 - Need For Surgery In Small Bowel Obstruction Can Be Suggested Based On A Deep Learning Analysis Of Diagnostic Computed Tomography

Image or Table (Do Not add text here):

Purpose/Background: Small bowel obstruction (SBO) is one of the most common causes of hospital admission in the U.S. with 1.8 million cases annually, resulting in 350,000 surgeries and $2.3 billion in healthcare costs each year. The two main treatment options for SBO are non-operative management and surgery. Choosing the appropriate modality for the patient depends on clinical, radiological, and laboratory analysis, along with the clinical acumen of the treating team. Nearly all patients presenting with clinical signs of SBO will undergo CT (computed tomography) scanning of the abdomen and pelvis. Previous studies have shown that imaging features such as bowel diameter and mesenteric edema are predictive of surgical management. To the best of our knowledge, deep learning (DL) model has not been used to predict the need for surgery in SBO. We hypothesize that a DL model based on CT scan data, could be a robust and objective means to guide human decisions.

Methods/Interventions: We retrospectively identified patients from 2013-22, aged 18 or older, having partial or complete SBO who underwent contrast-enhanced CT of the abdomen and pelvis. Patients with insufficient medical records or patients not admitted post-scan were excluded. Based on the treatment modalities patients were classified into two groups: operative and non-operative. Out of 4,113 CT scans of 3,203 patients meeting the inclusion criteria, 50 scans from each group were randomly selected, amounting to a total of 100 scans. We trained a convolutional neural network (CNN) to classify cases as requiring surgery or not. To address the small dataset size, we applied transfer learning by integrating a ResNet-18 feature extractor into our model, pre-trained on 23 medical datasets from the MedicalNet framework. This allowed the model to leverage prior knowledge from medical imaging tasks. Three additional convolutional layers were added to capture SBO surgery-specific features. To improve generalizability and prevent overfitting, we used normalization techniques and data augmentation, including scaling, rotation, and brightness adjustments. Our dataset is split into 60 CT images for training, 20 for validation and 20 for holdout testing, with CT volumes sized at 128 x 256 x 256 pixels. Basic demographics, including age and sex, were recorded, in addition to the AI model’s accuracy were evaluated by comparing its prediction to the actual clinical decision.

Results/Outcomes: The mean age of the 80 patients used for training and validation was 61.2 years (SD = 16.8). Out of 80 patients, 36 were male (45%) and 44 were female (55%). The model could classify cases requiring surgery vs non-surgery with 70% accuracy on validation set and 65% accuracy on the unseen held-out data, demonstrating its generalizability to unseen cases. We also evaluated the model's interpretability, as illustrated in Fig. 1, showing that it can identify relevant features in the bowel like dilation, reinforcing its reliability in surgical decision-making.

Conclusion/Discussion: In our early data set analysis, DL could predict the ultimate needed treatment option based on initial CT imaging with a high accuracy rate of 70% in our validation set and 65% in held-out test data. Despite the small dataset, the model demonstrated generalizability, and interpretability analysis confirmed its ability to identify relevant radiological features. Use of a DL model to guide treatment decisions is promising. We predict the accuracy of DL to predict outcomes to increase even further with further expansion analysis of our larger data set of over 4000 CT scans.