Automatic Hip Fracture Identification and Functional Subclassification\n with Deep Learning

Abstract

Purpose: Hip fractures are a common cause of morbidity and mortality.\nAutomatic identification and classification of hip fractures using deep\nlearning may improve outcomes by reducing diagnostic errors and decreasing time\nto operation. Methods: Hip and pelvic radiographs from 1118 studies were\nreviewed and 3034 hips were labeled via bounding boxes and classified as\nnormal, displaced femoral neck fracture, nondisplaced femoral neck fracture,\nintertrochanteric fracture, previous ORIF, or previous arthroplasty. A deep\nlearning-based object detection model was trained to automate the placement of\nthe bounding boxes. A Densely Connected Convolutional Neural Network (DenseNet)\nwas trained on a subset of the bounding box images, and its performance\nevaluated on a held out test set and by comparison on a 100-image subset to two\ngroups of human observers: fellowship-trained radiologists and orthopaedists,\nand senior residents in emergency medicine, radiology, and orthopaedics.\nResults: The binary accuracy for fracture of our model was 93.8% (95% CI,\n91.3-95.8%), with sensitivity of 92.7% (95% CI, 88.7-95.6%), and specificity\n95.0% (95% CI, 91.5-97.3%). Multiclass classification accuracy was 90.4% (95%\nCI, 87.4-92.9%). When compared to human observers, our model achieved at least\nexpert-level classification under all conditions. Additionally, when the model\nwas used as an aid, human performance improved, with aided resident performance\napproximating unaided fellowship-trained expert performance. Conclusions: Our\ndeep learning model identified and classified hip fractures with at least\nexpert-level accuracy, and when used as an aid improved human performance, with\naided resident performance approximating that of unaided fellowship-trained\nattendings.\n

Ähnliche Arbeiten