Development and validation of automated three-dimensional convolutional neural network model for acute appendicitis diagnosis

Figures

• 

  Fig. 1. Integration workflow of IA model. Information of Appendix (IA) represents a fully automated diagnostic framework employing a three-dimensional convolutional neural network (CNN). This model incorporates a pipeline featuring a two-stage binary algorithm connected to transfer learning, enabling the prediction of three classifications: non-appendicitis, simple appendicitis, and complicated appendicitis. DICOM files, sorted from the PACS platform, are pre-processed and anonymised. The Hounsfield unit scale of the extracted CT images is converted into 8-bit greyscale based on a window width of 60 and a level of 400. These grayscale images facilitate the localisation and generation of the volume of interest (VOI), which is then processed using the 3D-CNN algorithm. In the emergency medical system, the IA model simplifies the physician’s task to a single button push at the DICOM sorting stage, providing a diagnosis with associated probabilities for the input image within the VOI.

• 

  Fig. 2. Automated localisation for region of interest and generation of three-dimensional volume of interest for appendicitis. (a) In all cross-sectional slices of abdominopelvic computed tomography (APCT), automated localisation of the region of interest (ROI) on the z-axis, corresponding to the anatomical location of appendicitis, is performed using 2D-DenseNet169. The initial ROI range is defined by the highest level of the iliac crest and the lowest margin of the sacroiliac (SI) joint (indicated by the white bidirectional arrow) in the maximal intensity projection view (b). Extended upper and lower margins of the ROI (dotted line), with exploratory data analysis (EDA*), lead to the automated extraction of slices based on the extended z-axis ROI, with an additional upper margin of 50 mm referred to for the automated extraction of 3D-VOI. (c) Entire volume is cropped into the VOI using body segmentation and 3D-bounding box optimisation on the x- and y-axes. The refinement of the x- and y-axes of the VOI, achieved through body segmentation, is integrated with the z-axis to generate the initial 3D bounding box. Body segmentation that isolates only the patient area involves a sequence of…

• 

  Fig. 3. Automated detection of upper and lower margin for region of interest in abdominal computed tomography slices. (a) In a 43-year-old male patient with simple appendicitis, among all axial slices of the abdominal section (serial numbers 1 to 115), thirty slices (from serial numbers 52 to 81) were automatically sorted by DenseNet169 for the initial region of interest (ROI) range (blue box) with an extended upper margin (yellow bow). (b) Automated detection of the highest level of the iliac crest (white arrow) at slice serial number 62, and the lowest level of the sacroiliac (SI) joint at slice serial number 81 (yellow arrow), positioned between the iliac bone (red triangle) and sacrum (black asterisk). Bony structure (white arrow) is not visible in slice serial number 61 but is seen from slice serial number 62 onwards. Similarly, the iliac bony structure of the SI joint (yellow arrow) disappears after slice serial number 81.

• 

  Fig. 4. Exploratory data analysis for extended range of region of interest. EDA for 25, 50, and 75 mm-extended upper margins of the ROI on the z-axis using DenseNet169 indicated that the performance of each model was not significantly affected, as evidenced by consistent receiver operating characteristic (ROC) values. However, the model with a 50 mm threshold displayed the best performance (a). EDA for extended lower margins down to the upper border of the symphysis pubis, in comparison to model performance based on the SI joint, also displayed no significant difference in ROC values (b). (c) Model performance across three randomly split independent seeds was evaluated for the thresholds of 25, 50, and 75 mm-extended upper margins and the extended lower margin with an additional 50 mm from the lower margin of the sacroiliac joint. The analysis included area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score.

• 

  Fig. 5. Gradient-weighted class activation mapping for true classes. APCT (right) and gradient-weighted class activation mapping (Grad-CAM) images (left) of patients visiting the emergency room with right-sided lower quadrant pain are displayed in the internal validation dataset using DenseNet169. In the internal validation dataset, (a) dilated appendiceal lumen with the enhancement of a thickened wall (arrow) is shown alongside a Grad-CAM image indicative of subcecal type simple appendicitis, with a probability of 0.988. (b) Grad-CAM image corresponding to retrocecal type simple appendicitis (arrow) with a probability of 0.981. (c) Appendicolith (arrow) alongside a Grad-CAM image indicative of postileal-type simple appendicitis with a probability of 0.999. (d) Enhanced appendiceal lumen situated along the pelvic cavity (arrow) with a corresponding Grad-CAM image for pelvic type simple appendicitis, showing a probability of 0.991. (e) Multi-loculated cavity surrounding mural haziness of the appendix (arrow) and a Grad-CAM image indicative of complicated appendicitis with a probability of 0.928. In the external validation dataset, Grad-CAM images (f–i) displayed heatmaps correlated to the…