A scalable physician-level deep learning algorithm detects universal trauma on pelvic radiographs

Abstract

Pelvic radiograph (PXR) is essential for detecting proximal femur and pelvis injuries in trauma patients, which is also the key component for trauma survey. None of the currently available algorithms can accurately detect all kinds of trauma-related radiographic findings on PXRs. Here, we show a universal algorithm can detect most types of trauma-related radiographic findings on PXRs. We develop a multiscale deep learning algorithm called PelviXNet trained with 5204 PXRs with weakly supervised point annotation. PelviXNet yields an area under the receiver operating characteristic curve (AUROC) of 0.973 (95% CI, 0.960-0.983) and an area under the precision-recall curve (AUPRC) of 0.963 (95% CI, 0.948-0.974) in the clinical population test set of 1888 PXRs. The accuracy, sensitivity, and specificity at the cutoff value are 0.924 (95% CI, 0.912-0.936), 0.908 (95% CI, 0.885-0.908), and 0.932 (95% CI, 0.919-0.946), respectively. PelviXNet demonstrates comparable performance with radiologists and orthopedics in detecting pelvic and hip fractures.

Fig. 1. The source and distribution of each data set.

We included 5204 pelvic radiographs (PXR) from 2008 to 2016 as our development data set to develop PelviNet. Then we included PXRs from 1888 patients presented in the emergency room from January to December 2017 as our test data set. In advance, we randomized selected 150 PXRs from the test dataset to compose the PXR150 data set, which was used to perform the physician comparison test.

Fig. 2. The receiver operating characteristic curve and precision-recall curve of the universal trauma finding detection algorithm.

a The receiver operating characteristic (ROC) curve of the performance of PelviXNet in the clinical scenario and the cross mark represents the performance on the probability cutoff value. b The precision-recall (PR) curve of the performance of PelviXNet and the cross mark represents the performance on the probability cutoff value. The 95% confidence intervals (CIs) of the ROC and PR curves were estimated using bootstrapping with 2000 replicates, which was indicated as the purple area in the panels.

Fig. 3. The illustration of heatmap overlaid by the algorithm on original images.

The red color represents a high probability of acute trauma finding detected. The heatmap indicates a no fracture. b Anterior–posterior compression type pelvic fracture, c left femoral non-displaced intertrochanteric fracture, d left periprosthetic fracture, e right femoral shaft fracture, and f right hip dislocation. g Difficult case of pelvic fracture. h Clinically missed pelvic fracture. i Multiple pelvic fractures detected simultaneously by PelviXNet, respectively. All the pelvic fracture examples except case h in this figure require angioembolization due to massive hemorrhage.

Fig. 4. The illustration of bone fracture classification and localization system.

The proposed bone fracture classification and localization system trained with point-based supervision signals. The network uses PXR images as input and extracts different levels of feature abstractions through a bottom-up pathway, which is further fused by a top-down path.

Fig. 5. The Overviw of the ensemble method for inference in testing.

During the inference stage, once the image was input, five models with five augmented images each were ensembled to generate the final prediction of the image.