Automated classification of hip fractures using deep convolutional neural networks with orthopedic surgeon-level accuracy: ensemble decision-making with antero-posterior and lateral radiographs

Abstract

Background and purpose — Deep-learning approaches based on convolutional neural networks (CNNs) are gaining interest in the medical imaging field. We evaluated the diagnostic performance of a CNN to discriminate femoral neck fractures, trochanteric fractures, and non-fracture using antero-posterior (AP) and lateral hip radiographs. Patients and methods — 1,703 plain hip AP radiographs and 1,220 plain hip lateral radiographs were included in the total dataset. 150 images each of the AP and lateral views were separated out and the remainder of the dataset was used for training. The CNN made the diagnosis based on: (1) AP radiographs alone, (2) lateral radiographs alone, or (3) both AP and lateral radiographs combined. The diagnostic performance of the CNN was measured by the accuracy, recall, precision, and F1 score. We further compared the CNN’s performance with that of orthopedic surgeons. Results — The average accuracy, recall, precision, and F1 score of the CNN based on both anteroposterior and lateral radiographs were 0.98, 0.98, 0.98, and 0.98, respectively. The accuracy of the CNN was comparable to, or statistically significantly better than, that of the orthopedic surgeons regardless of radiographic view used. In the CNN model, the accuracy of the diagnosis based on both views was significantly better than the lateral view alone and tended to be better than the AP view alone. Interpretation — The CNN exhibited comparable or superior performance to that of orthopedic surgeons to discriminate femoral neck fractures, trochanteric fractures, and non-fracture using both AP and lateral hip radiographs.

Machine Learning Solutions for Osteoporosis—A Review

“…( 77 ) Twelve studies compared ML model performance versus human experts. ( 66,71,78–83,85,91,93,94 ) In four of these studies, ML outperformed human experts significantly. ( 80,83,85,91 ) Thirteen studies applied transfer learning based on pre‐defined CNN architectures, pre‐trained on the ImageNet data set ( 77,79–83,85,90,91,93,95 ) or on a radiography image database.…”

“…( 75,76,90–97 ) Nineteen studies developed CNN models for image analysis. ( 66,67,71,72,77–86,90,91,93–95 ) Others used features extracted from images or collected from non‐imaging data. ( 68–70,73–77,87–89,92,96,97 ) Studies reported average best AUC of 0.92 (range 0.63 to 1.00) and average best accuracy of 89.8% (range 78.4% to 99.1%].…”

“…Among the 32 studies that investigated fracture detection (Table 3), (66–97 ) 11 were on vertebral fractures, ( 66–76 ) 17 hip fractures, ( 74–90 ) and 10 other fracture sites such as humerus or wrist. ( 75,76,90–97 ) Nineteen studies developed CNN models for image analysis.…”

“…( 78,94 ) Some of these were fine‐tuned on the last layers, ( 82 ) on the penultimate layer, (77,79,83,93,95 ) or were unclear. ( 78,80,81,85,90,91,94 ) Model decision‐making was highlighted using heat maps such as Grad‐CAM. ( 66,67,78,79,84,90,94 ) However, several studies did not use techniques that allowed for adequate understanding of the models' decision‐making (eg, data visualization).…”

“…( 29,51,60,72,92 ) Acknowledging potential biases or limitations is fundamental for the accurate interpretation and especially for claims related to clinical implication. Nevertheless, several studies showed a greater awareness of the potential existence of bias and made efforts to address this either directly by controlling for the risks of overfitting, ( 23,28,30,32,34,37–40,43–46,49,53,54,56,57,59,64,77–85,90,91,93–95 ) indirectly by visualizing decision‐making, ( 36,38,48,66,67,78,79,84,90,94,104 ) or by using gold standard strategies for comparison. ( 66,71,78–83,85,91,93,94,104,107 )…”