Deep Learning for the Differential Diagnosis between Transient Osteoporosis and Avascular Necrosis of the Hip

Michail E Klontzas; Ioannis Stathis; Konstantinos Spanakis; Aristeidis H Zibis; Kostas Marias; Apostolos H Karantanas

doi:10.3390/diagnostics12081870

Deep Learning for the Differential Diagnosis between Transient Osteoporosis and Avascular Necrosis of the Hip

Diagnostics (Basel). 2022 Aug 2;12(8):1870. doi: 10.3390/diagnostics12081870.

Authors

Michail E Klontzas^{1

2

3

4}, Ioannis Stathis¹, Konstantinos Spanakis¹, Aristeidis H Zibis⁵, Kostas Marias^{2

3

6}, Apostolos H Karantanas^{1

2

3

4}

Affiliations

¹ Department of Medical Imaging, University Hospital, 71110 Heraklion, Greece.
² Computational BioMedicine Laboratory, Institute of Computer Science, Foundation for Research and Technology (FORTH), 70013 Heraklion, Greece.
³ Advanced Hybrid Imaging Systems, Institute of Computer Science, Foundation for Research and Technology (FORTH), 70013 Heraklion, Greece.
⁴ Department of Radiology, School of Medicine, University of Crete, Voutes Campus, 71003 Heraklion, Greece.
⁵ Department of Anatomy, Medical School, University of Thessaly, 41334 Larissa, Greece.
⁶ Department of Electrical & Computer Engineering, Hellenic Mediterranean University, 71004 Heraklion, Greece.

Abstract

Differential diagnosis between avascular necrosis (AVN) and transient osteoporosis of the hip (TOH) can be complicated even for experienced MSK radiologists. Our study attempted to use MR images in order to develop a deep learning methodology with the use of transfer learning and a convolutional neural network (CNN) ensemble, for the accurate differentiation between the two diseases. An augmented dataset of 210 hips with TOH and 210 hips with AVN was used to finetune three ImageNet-trained CNNs (VGG-16, InceptionResNetV2, and InceptionV3). An ensemble decision was reached in a hard-voting manner by selecting the outcome voted by at least two of the CNNs. Inception-ResNet-V2 achieved the highest AUC (97.62%) similar to the model ensemble, followed by InceptionV3 (AUC of 96.82%) and VGG-16 (AUC 96.03%). Precision for the diagnosis of AVN and recall for the detection of TOH were higher in the model ensemble compared to Inception-ResNet-V2. Ensemble performance was significantly higher than that of an MSK radiologist and a fellow (P < 0.001). Deep learning was highly successful in distinguishing TOH from AVN, with a potential to aid treatment decisions and lead to the avoidance of unnecessary surgery.

Keywords: Artificial Intelligence; Inception-ResNetV2; InceptionV3; MR imaging; VGG-16; avascular necrosis; deep learning; hip; osteoporosis/transient; transfer learning.

Grants and funding

This research received no external funding.