Convolutional-neural-network-based diagnosis of appendicitis via CT scans in patients with acute abdominal pain presenting in the emergency department

Figures

• 

  Figure 1. Schematic overview of a deep convolutional neural network approach for the diagnosis of acute appendicitis (upper) and the validation process (lower). For internal validation, 8-fold cross validation was used. The dataset was randomly separated into 8 datasets. Hyperparameters such as the learning rate, the number of epochs, and the number of layers were determined during the first model training and then used to train the other 7 models. The remaining 1/8 of the data were used for a test. For external validation, a CT dataset acquired from two institutions was used for the 8 differently trained CNN models. The final classification of acute appendicitis through a CNN was determined using the average of the network outputs [drawn by ITK-SNAP (version 3.6; http://www.itksnap.org/pmwiki/pmwiki.php) and Microsoft PowerPoint].

• 

  Figure 2. Confusion matrix (upper) used to diagnose acute appendicitis in the trained CNN using datasets for each of the three institutions and ROC curve analysis (lower) performed to diagnose acute appendicitis from a trained CNN using datasets from each of the three institutions considered (drawn by Microsoft PowerPoint and Matlab 2018b).

• 

  Figure 3. A 44-year-old woman who visited the ER with abdominal pain and a normal appendix. Original CT images within extracted 3D isotropic cubes show a normal appendix as an air-filled tubular structure (left side). Corresponding heatmap overlay using Grad-CAM highlights a normal appendix that is correctly recognized by the trained CNN (true negative) (right side).

• 

  Figure 4. A 62-year-old woman who visited the ER with abdominal pain and was diagnosed with acute appendicitis. Original CT images within the extracted 3D isotropic cubes show a dilated appendix with wall thickening and appendicolith, compatible with acute appendicitis (left side). An inflamed appendix is highlighted on the heatmap overlay using Grad-CAM, which means the algorithm detected the appendix properly and diagnosed acute appendicitis accurately by assigning a given weight to the image location corresponding to the appendix (true positive) (right side).

• 

  Figure 5. A 54-year-old woman who visited the ER with abdominal pain and was diagnosed with acute appendicitis. Original CT images within the extracted 3D isotropic cubes show an inflamed appendix with wall thickening and appendicolith, compatible with acute appendicitis (arrowhead) (left side). Corresponding heatmap overlay obtained using Grad-CAM showing that the terminal ileum (arrow) is incorrectly recognized by the trained CNN as a normal appendix (false negative) (right side).

• 

  Figure 6. A 48-year-old man who visited the ER with abdominal pain and was diagnosed with cecal diverticulitis. Original CT images within the extracted 3D isotropic cubes show thick-walled diverticulum (arrow) in the cecum, fat stranding and peritoneal thickening for a normal appendix (arrowhead) (left side). Corresponding heatmap overlay obtained using Grad-CAM that represents an inflamed cecal diverticulum (arrow) that was recognized by the CNN as an inflamed appendix (false positive) (right side).