Artificial intelligence models assisting physicians in quantifying pancreatic necrosis in acute pancreatitis

Background: Acute pancreatitis (AP) is a potentially life-threatening condition characterized by inflammation of the pancreas, which can lead to complications such as pancreatic necrosis. The modified computed tomography severity index (MCTSI) is a widely used tool for assessing the severity of AP, particularly the extent of pancreatic necrosis. The accurate and timely assessment of the necrosis volume is crucial in guiding treatment decisions and improving patient outcomes. However, the current diagnostic process relies heavily on the manual interpretation of computed tomography (CT) scans, which can be subjective and prone to variability among clinicians. This study aimed to develop a deep-learning network model to assist clinicians in diagnosing the volume ratio of pancreatic necrosis based on the MCTSI for AP.

Methods: The datasets comprised retrospectively collected plain and contrast-enhanced CT scans from 144 patients (6 with scores of 0 points, 42 with scores of 2 points, and 65 with scores of 4 points) and the National Institutes of Health contrast-enhanced CT scans from 45 patients with scores of 0 points. An improved fully convolutional neural networks for volumetric medical image segmentation (V-Net) model was developed to segment the pancreatic volume (i.e., the whole pancreas, necrotic pancreatic tissue, and non-necrotic pancreatic tissue) and to quantify the split volume ratios. The improved strategy included three stages of body up- and down-sampling adapted to the task of segmentation in AP, and the selection of objects, loss function, and smoothing coefficients. The model interpretations were compared with those of clinicians with different levels of experience. The reference standard was manually segmented by a pancreatic radiologist. Accuracy, macro recall, and macro specificity were employed to compare the diagnostic efficacy of the model and the clinicians.

Results: In total, 144 patients (mean age: 44±13 years; 40 females, 104 males) were included in the study. Optimal training results were obtained using the necrotic pancreatic tissue and whole pancreas as the input objects, and combining dice loss and 500 smoothing coefficients as the loss function for training. The dice coefficient for the whole pancreas was 0.811 and that for the necrotic pancreatic tissue was 0.761. The performance of the artificial intelligence model and clinicians were compared. The accuracy, macro recall, and macro specificity of the improved V-net were 0.854, 0.850 and 0.923, respectively, which were all significantly higher than those of the senior and junior clinicians (P<0.05).

Conclusions: Our proposed model could improve the effectiveness of clinicians in diagnosing pancreatic necrosis volume ratios in clinical settings.