Premium
Computer vs human: Deep learning versus perceptual training for the detection of neck of femur fractures
Author(s) -
Adams Matthew,
Chen Weijia,
Holcdorf David,
McCusker Mark W,
Howe Piers DL,
Gaillard Frank
Publication year - 2019
Publication title -
journal of medical imaging and radiation oncology
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.31
H-Index - 43
eISSN - 1754-9485
pISSN - 1754-9477
DOI - 10.1111/1754-9485.12828
Subject(s) - convolutional neural network , medicine , artificial intelligence , perception , subspecialty , deep learning , perceptual learning , radiography , radiology , pattern recognition (psychology) , computer science , pathology , psychology , neuroscience
To evaluate the accuracy of deep convolutional neural networks ( DCNN s) for detecting neck of femur (NoF) fractures on radiographs, in comparison with perceptual training in medically‐naïve individuals. Methods This study extends a previous study that conducted perceptual training in medically‐naïve individuals for the detection of NoF fractures on a variety of dataset sizes. The same anteroposterior hip radiograph dataset was used to train two DCNN s (AlexNet and GoogLeNet) to detect NoF fractures. For direct comparison with perceptual training results, deep learning was completed across a variety of dataset sizes (200, 320 and 640 images) with images split into training (80%) and validation (20%). An additional 160 images were used as the final test set. Multiple pre‐processing and augmentation techniques were utilised. Results AlexNet and GoogLeNet DCNN s NoF fracture detection accuracy increased with larger training dataset sizes and mildly with augmentation. Accuracy increased from 81.9% and 88.1% to 89.4% and 94.4% for AlexNet and GoogLeNet respectively. Similarly, the test accuracy for the perceptual training in top‐performing medically‐naïve individuals increased from 87.6% to 90.5% when trained on 640 images compared with 200 images. Conclusions Single detection tasks in radiology are commonly used in DCNN research with their results often used to make broader claims about machine learning being able to perform as well as subspecialty radiologists. This study suggests that as impressive as recognising fractures is for a DCNN , similar learning can be achieved by top‐performing medically‐naïve humans with less than 1 hour of perceptual training.