E5076. Novel Convolutional Neural Network Based Automated Rib Fracture Detection and Classification in the Acute Trauma Patient and Beyond
  1. Bilal Quadri; Santa Clara Valley Medical Center
  2. Randol Spaulding; Santa Clara University
  3. Michelle Lam; Santa Clara Valley Medical Center
  4. Mahesh Patel; Santa Clara Valley Medical Center
  5. Benjamin Chou; Santa Clara Valley Medical Center
  6. Yuling Yan; Santa Clara University
  7. Young Kang; Santa Clara Valley Medical Center
To develop a robust and universal convolutional neural network (CNN) that automatically and accurately identifies acute rib fractures on CT scans in real-time trauma setting, offering accuracy equal to or exceeding that of a radiologist and existing algorithms by utilizing a novel single point-based labeling method to capture complex 3D fractures.

Materials and Methods:
CT studies for trauma patients at a county hospital level 1 trauma center from 2017–2022 with CT reports positive for rib fracture were extracted from the institutional PACS and anonymized. Rib fractures on images were labeled using VGG Image Annotator (VIA) software by placing a single point in the estimated center of the fracture (fracture centroid). Labeling of the fractures on CT was performed by two radiologists. This was supplemented with existing dataset from the Medical Image Computing and Computer Assisted Interventions (MICCAI) conference 2020 RibFrac Challenge. The entire data set was divided into training, validation, and test sets. Each fracture centroid was converted into a 3D continuous label. A novel loss function was developed to train a CNN on continuous labels. This was incorporated with a CNN with U-Net architecture developed for segmentation of fracture centroids. The CNN was then trained on random subvolumes of each CT scan. A 2D CNN was used to extract and compile a cumulative segmentation map from each individual slice. Fractures were then detected by analyzing the precision-recall curve over the sampled subvolumes in the validation set and evaluated on the test set.

Our database yielded 206 unique CTs containing 4944 ribs and 2781 rib fractures. There were 626 CT studies obtained from the 2020 RibFrac Challenge. The number of rib fractures in the training, validation, and testing sets were 5130 (87.4%), 198 (3.4%), and 540 (9.2%), respectively. The algorithm detected rib fractures with a sensitivity of 85.4% (p < 0.0001), with an average of 0.9 false positives per scan.

CNN utilizing a novel method of automated point-based labeling achieved a promising fracture detection sensitivity of 85.4% (p < 0.0001), comparable to traditional, nonautomated approaches. An algorithm trained with a data set employing this labeling method simplifies data preparation. We have demonstrated a method potentially facilitating continued development and refinement of algorithms to aid the radiologist in detecting rib fractures in patients presenting with trauma or other chest symptoms. In the future, point-based labeling can be utilized for other nonuniformly shaped regions of interest in cross-sectional imaging beyond the ribs and chest.