Automated Fracture Detection from CT Scans

Computed Tomography (CT) scans play a crucial role in modern medical imaging for detecting bone fractures. However, identifying the location and position of broken bones can be challenging, particularly in complex cases involving multiple extremities. In this paper, we propose a robust approach for...

Full description

Saved in:
Bibliographic Details
Published in2023 IEEE Conference on Artificial Intelligence (CAI) pp. 161 - 162
Main Authors Chaudhary, Abhishek Kumar, Roy, Shotabdi, Rizk, Rodrigue, Santosh, KC
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.06.2023
Subjects
Bones
Computed tomography
CT scans
Fractured Bones
Location awareness
Neural networks
Pre-Trained Neural Network Model
Predictive models
Scalability
Training
YOLO v7
Online AccessGet full text

Cover

More Information
Summary:Computed Tomography (CT) scans play a crucial role in modern medical imaging for detecting bone fractures. However, identifying the location and position of broken bones can be challenging, particularly in complex cases involving multiple extremities. In this paper, we propose a robust approach for enhancing fracture detection and localization in CT scans using the YOLO v7 model. By simultaneously predicting class probabilities and bounding boxes in a single iteration, the YOLO v7 model shows improved and consistent performance measures. We developed our approach on a dataset of 1217 CT cases, by training our model on combined extremities, resulting in improved and consistent performance metrics for detecting and localizing fractures. Our proposed method achieved a high precision rate of 99% for identifying broken bones in the lower right limb and 66% for the combined set of upper and lower extremities on both sides. Our findings highlight the potential of YOLO v7 as a powerful tool for enhancing medical imaging workflows, particularly for further treatment planning, by improving fracture detection and localization. Future studies could investigate the generalizability and scalability of our proposed method in larger datasets and different clinical settings.
DOI:10.1109/CAI54212.2023.00077