PaperlayerView exact revision on arXiv

Source-linked AI summary

Fracture Detection in Wrist X-ray Images Using Deep Learning-Based Object Detection Models

Fırat Hardalaç, Fatih Uysal, Ozan Peker, Murat Çiçeklidağ, Tolga Tolunay, Nil Tokgöz, Uğurhan Kutbay, Boran Demirciler, Fatih Mert

arXiv:2111.07355v3eess.IVcs.CVcs.LGphysics.med-ph

TL;DR

The study evaluates deep-learning detection of wrist fractures in X-rays collected at Gazi University Hospital. Its WFD-C ensemble reached 0.8639 AP50, above the best pre-ensemble PAA result.

  • Problem

    Wrist-fracture detection from X-rays is the study’s focus within a diagnostic process that also uses medical history and physical examination.

    Source: §1 · p. 2

  • Method

    The study compared 10 object-detection models across 20 procedures and developed five ensemble models, combined into WFD-C.

    Source: §1 · p. 3

  • Results

    0.8639 AP50 in WFD-C exceeded the best pre-ensemble result, 0.754 AP50 in PAA, by over 10%.

    Source: §4.4.3 · p. 22

  • Takeaways & Limitations

    The authors say the clinical dataset gives the ensemble potential for future hospital use and the augmentation method may guide similar medical X-ray studies.

    Source: §5 · p. 27

    Fractures of the small carpal bones were not studied and were ignored in labeling.

    Source: §5 · p. 28

Abstract

from arXiv · show

Hospitals, especially their emergency services, receive a high number of wrist fracture cases. For correct diagnosis and proper treatment of these, images obtained from various medical equipment must be viewed by physicians, along with the patients medical records and physical examination. The aim of this study is to perform fracture detection by use of deep learning on wrist Xray images to support physicians in the diagnosis of these fractures, particularly in the emergency services. Using SABL, RegNet, RetinaNet, PAA, Libra R_CNN, FSAF, Faster R_CNN, Dynamic R_CNN and DCN deep learning based object detection models with various backbones, 20 different fracture detection procedures were performed on Gazi University Hospitals dataset of wrist Xray images. To further improve these procedures, five different ensemble models were developed and then used to reform an ensemble model to develop a unique detection model, wrist fracture detection_combo (WFD_C). From 26 different models for fracture detection, the highest detection result obtained was 0.8639 average precision (AP50) in the WFD-C model. Huawei Turkey R&D Center supports this study within the scope of the ongoing cooperation project coded 071813 between Gazi University, Huawei and Medskor. Code is available at https://github.com/fatihuysal88/wrist-d

1. Introduction

The introduction motivates deep-learning detection of wrist fractures in clinical X-rays and outlines a program of model comparisons, augmentation, and ensemble development.

1. Introduction

  • Wrist X-rays from Gazi University Hospital supplied the clinical dataset for the detection work.The introduction notes that X-rays are preferred over MRI and CT because they are more cost-efficient.Source: §1 · p. 2
  • The fracture focus was the radius and ulna, beneath the wrist’s eight carpal bones.The introduction names the eight carpal bones and describes their connection between hand and forearm.Source: §1 · p. 2
  • The proposed WFD-C ensemble was reported to improve on the best literature AP result by approximately 10%.The study also introduces LRP evaluation for biomedical bone-fracture detection.Source: §1 · p. 3
  • The study compared 10 deep-learning models across 20 procedures, testing augmentation methods and developing five ensemble models.The authors identify an augmentation method they found most compatible with X-ray images.Source: §1 · p. 3
  • The study used AP, AR, and LRP to evaluate fracture detection, reporting LRP use in biomedical bone-fracture evaluation for the first time.Source: §1 · p. 3

2. Related Works

Related work covers fracture detection across anatomical sites and imaging modalities, alongside separate studies classifying fracture types; the authors position their work as wrist-X-ray detection with multiple detectors and ensembles.

2. Related Works

  • Prior detection reports include 62.04% AP for arm X-rays, 77.4% AP for mixed-region radiographs, and 92.9% sensitivity for rib CT.Source: §2 · pp. 3–4
  • Separate classification work predicts fracture class, with reported examples including 99.1% accuracy on thoracolumbar CT and 0.954 AUC on wrist radiographs.Source: §2 · p. 4
  • Prior detection commonly used Faster R-CNN or YOLO with ResNet or ResNeXt backbones; this study applies nine detector families to wrist X-rays and develops ensembles.Source: §2 · p. 4

3. Methods

The methods apply transfer learning to object detectors for wrist-fracture localization, compare procedures with and without augmentation, and use their results to develop ensemble models.

3. Methods

  • The study performed 20 wrist-fracture detection procedures with and without augmentation across the listed detector families and backbone configurations.The methods include DCN, Dynamic R-CNN, Faster R-CNN, FSAF, Libra R-CNN, PAA, RetinaNet, RegNet, and SABL.Source: §3 · p. 5
  • All detectors used transfer learning from COCO-pretrained weights, with the object categories reduced from 80 to one fracture category.Figure 2 presents the detection models as using transfer learning.Source: §3 · p. 5, Figure 2 · p. 5
  • The authors developed new ensemble models based on the detection results to further improve wrist-fracture detection.Source: §3 · p. 5

3.1. Fracture Detection Models Based on DL for Wrist X-ray Images

The section introduces single- and two-stage wrist-fracture detectors and contrasts their architectures and typical speed–accuracy trade-offs. It also compares PAA and Dynamic R-CNN bounding-box outputs across thresholds.

3.1. Fracture Detection Models Based on DL for Wrist X-ray Images

  • Single-stage detectors pass images through a backbone, neck, and DenseHead to output object locations and classes; two-stage detectors add an RoIHead.Figures 3 and 4 depict the respective detector types.Source: §3.1 · p. 6, Figure 3 · p. 6, Figure 4 · p. 6
  • Single-stage detectors are generally faster to train, while two-stage detectors generally have higher accuracy, though the pattern may vary by dataset.Source: §3.1 · p. 6
  • PAA produces many bounding boxes at low probability values, whereas Dynamic R-CNN produces fewer outputs at higher probability values across thresholds from 0.1 to 0.9.Table 1 and Figure 5 present the bbox counts by threshold.Source: §3.1 · p. 7, Table 1 · p. 7, Figure 5 · p. 7
  • Single-stage models used include FSAF, PAA, RetinaNet and RetinaNet-based SABL, Libra, and RegNet; two-stage models include Dynamic R-CNN, Faster R-CNN, SABL, and DCN.The study used FPN in the neck and ResNet50 as the backbone where applicable, with RegNetX-3.2GF also used for RegNet.Source: §3.1 · pp. 6–7

3.1.1. RetinaNet

The study uses a COCO-pretrained RetinaNet with a ResNet50 backbone for wrist-fracture detection.

3.1.1. RetinaNet

  • The study uses a RetinaNet object detector pretrained with COCO and a ResNet50 backbone for wrist-fracture detection.Source: §3.1.1 · p. 7
  • RetinaNet is a single-stage object detector that the passage says can achieve higher detection results than many two-stage detectors.Source: §3.1.1 · p. 7
  • RetinaNet comprises a ResNet backbone, an FPN neck, and a RetinaHead with classification and box-regression subnets.Source: §3.1.1 · p. 7

3.1.2. Feature Selective Anchor-Free (FSAF)

FSAF is described as an anchor-free model for single-shot object detectors and is RetinaNet-based.

3.1.2. Feature Selective Anchor-Free (FSAF)

  • FSAF is an anchor-free model for single-shot object detectors.Source: §3.1.2 · p. 8
  • The FSAF models examined are RetinaNet-based, with ResNet as backbone and FPN in the neck.Source: §3.1.2 · p. 8
  • For FSAF in this study, FSAFHead is used as the DenseHead instead of RetinaHead.Source: §3.1.2 · p. 8

3.1.3. Dynamic R-CNN

The study uses COCO-pretrained Dynamic R-CNN with a ResNet50 backbone for wrist-fracture detection.

3.1.3. Dynamic R-CNN

  • Dynamic R-CNN combines Dynamic Label Assignment (DLA) and Dynamic SmoothL1 Loss (DSL).Source: §3.1.3 · p. 8
  • DLA dynamically increases the IoU threshold to improve proposal quality, while DSL adjusts the regression-loss shape to support that improvement.Source: §3.1.3 · p. 8
  • The study uses Dynamic R-CNN pretrained on COCO with a ResNet50 backbone for fracture detection.Source: §3.1.3 · p. 8

3.1.4. Libra R-CNN

Libra R-CNN uses balancing components to improve detection, while Faster R-CNN uses a Region Proposal Network for region proposals. The study used Libra RetinaNet and Faster R-CNN with a ResNet50 backbone.

3.1.4. Libra R-CNN

  • Libra R-CNN combines IoU-balanced sampling, a balanced feature pyramid, and balanced L1 loss to reduce imbalances at sample, feature, and objective levels.Source: §3.1.4 · p. 8
  • The study used the Libra RetinaNet model for wrist-fracture detection, as allowed by its hardware.Source: §3.1.4 · p. 8

3.1.5. Faster R-CNN

  • Faster R-CNN uses a Region Proposal Network to find region proposals and was used with a ResNet50 backbone for wrist-fracture detection.Source: §3.1.5 · p. 8

3.1.6. Side-Aware Boundary Localization (SABL)

The study applied SABL, DCNv2, and PAA to wrist-fracture detection using specified detector and backbone configurations. Their methods respectively localize box sides, model transformations, and assign anchors by fitted score probabilities.

3.1.6. Side-Aware Boundary Localization (SABL)

  • SABL localizes each bounding-box side through a dedicated network branch and was applied to Faster R-CNN and RetinaNet with ResNet50 backbones.Source: §3.1.6 · p. 8

3.1.7. Deformable Convolutional Networks (DCN)

  • DCNv2 combines deformable convolution and deformable RoI pooling to improve CNN transformation modeling, and the study used it with Faster R-CNN and ResNet50.Source: §3.1.7 · p. 8

3.1.8. Probabilistic Anchor Assignment (PAA)

  • PAA fits probability distributions to anchor scores, then trains on positive and negative samples assigned by their probabilities; the study used a ResNet50-backed version.Source: §3.1.8 · pp. 8–9

3.1.9. RegNet

RegNet determines stage widths and depths with a quantized linear function and parameterizes network populations. In this study, hardware constraints limited its fracture-detection use to RetinaNet with a RegNetX-3.2GF backbone.

3.1.9. RegNet

  • RegNet determines network stage widths and depths using a quantized linear function.Source: §3.1.9 · p. 9
  • Its design paradigm parameterizes network populations and combines advantages of manual design and neural architecture search.Source: §3.1.9 · p. 9
  • Hardware constraints limited the study to RegNet RetinaNet with a RegNetX-3.2GF backbone for fracture detection.Source: §3.1.9 · p. 9

3.2. Proposed Fracture Detection Models Based on DL for Wrist X-ray Images

The study built five WFD ensembles using weighted boxes fusion (WBF), selecting submodels according to detector structure or AP50, AR, and LRP criteria. It then combined those five ensembles into WFD-C.

3.2. Proposed Fracture Detection Models Based on DL for Wrist X-ray Images

  • Five WFD models were developed to improve fracture-detection results, using submodels selected by threshold criteria and detector structure.Source: §3.2 · p. 9, Wrist Fracture Detection (WFD)… · p. 9
  • WFD-1 and WFD-2 select submodels by single-stage and two-stage detector structure, respectively.Their most compatible model and weight combinations were selected from 40,593 single-stage and 388 two-stage combinations.Source: Wrist Fracture Detection (WFD)… · p. 9
  • WFD-3, WFD-4, and WFD-5 select submodels by optimal AP50, AR, and LRP thresholds, respectively.WFD-3 uses the five best AP50 models; WFD-4 uses the five best AR models, and WFD-5 uses DCN, Faster R-CNN, FSAF, Libra RetinaNet, and RetinaNet with the best LRP thresholds.Source: Wrist Fracture Detection (WFD)… · p. 9
  • For each WFD, the authors tried submodel combinations from two detectors up to the available maximum and varied integer weights across the detector range.WFD-1, for example, draws from six single-stage models with weights from 1 to 6.Source: Wrist Fracture Detection (WFD)… · pp. 9–10
  • WBF uses bounding-box confidence scores to produce averaged boxes, and the study combined detector outputs using varying weight coefficients.The workflow converts MMDetection JSON outputs to CSV before fusion.Source: Wrist Fracture Detection (WFD)… · pp. 10–11
  • WFD-C applies another WBF ensemble procedure to the previously developed WFD-1 through WFD-5 models.The scheme is shown in Figure 8, and the submodels are used in six ensemble models in total; their weight coefficients are listed in Table 2.Source: Wrist Fracture Detection (WFD)… · p. 11, Figure 8 · p. 11, Table 2 · p. 11

4. Experiments

The experiments evaluated wrist-fracture detection on Gazi University Hospital X-rays using deep-learning procedures and ensembles, including comparisons of WFD-C with individual models and across test-set sizes and predicted-box performance.

4.4. Fracture Detection Results

  • Using wrist X-rays from a Gazi University Hospital dataset, the study performed 20 different deep-learning object-detection procedures with models featuring varied backbones, using MMDetection for detection.All procedures used 40 epochs, SGD with momentum 0.9, and a learning rate of 0.001 that decreased 10 times at epochs 5, 10, 15, 25 and 35. Augmentation steps and other optimization settings were the same across models.Source: §4 · p. 11, §4.4 · pp. 15–16

4.1. Dataset and Labeling of Wrist X-ray Images

  • The fracture-positive wrist X-ray dataset contained 542 images from 275 emergency-department patients, collected during 2010–2020 on one Samsung GC70 machine, with patient-separated splits of 434 training, 54 validation, and 54 test images.Physicians labeled fracture areas belonging only to the radius and ulna.Source: §4.1 · pp. 12–13

4.2. Data Preprocessing of Wrist X-ray Images

  • Images were manually cropped to include the hand and wrist, resized to 800 × 800 × 3, and standardized using binary K-means background detection, inversion when needed, and CLAHE.CLAHE used an 11 × 11 grid and a 7.0 clip limit.Source: §4.2 · p. 14

4.3. Data Augmentation of Wrist X-ray Images

  • Random brightness-contrast augmentation contributed most to detection among the tested Albumentations methods and was used alongside random flipping at ratio 0.5.The tested methods also included sharpness, noise, gamma, Gaussian blur, and median blur, individually or in combination.Source: §4.3 · pp. 14–15

4.4.1. Evaluation Metrics

  • Evaluation used AP, AR, precision-recall curves, image-level ground-truth versus prediction comparisons, and LRP error; IoU thresholds determine true versus false detections.The passage defines AP as average precision across unique recall levels and AR across IoU values from 0.5 to 1.0.Source: §4.4.1 · p. 16, Table 4 · p. 16

4.4.2. Fracture Detection Results of 20 DL-Based Models

  • Dynamic R-CNN variants achieved the best validation AP50, while FSAF had the lowest validation LRP optimal threshold among the models.The epoch of highest validation AP50 ranged from 6–21 without augmentation and 5–12 with augmentation; SABL Faster R-CNN had the lowest train bbox loss, and DCN Faster R-CNN the lowest train loss.Source: Figure 13 · p. 17, Figure 14 · p. 17, Table 6 · p. 18
  • 0.754 AP50 was the highest test score among the individual models, achieved by PAA with augmentation.Among augmented models, the lowest LRP optimal threshold, oLRPLoc, oLRPFP, oLRPFN, and oLRP were reported for FSAF, SABL Faster R-CNN, RegNet RetinaNet, PAA/RetinaNet, and PAA, respectively.Source: Table 8 · p. 20, §4.4.2 · p. 20, Table 9 · p. 20

4.4.3. Fracture Detection Results of Proposed Models

  • The WFD-C ensemble achieved the highest AP50 score, 0.8639, compared with 0.754 for the PAA model, the highest pre-ensemble detection result; the authors report an increase of over 10%.Source: §4.4.3 · p. 22
  • Across 26 detection models, including six ensembles, WFD-C had the predicted boxes closest to ground truth and the fewest predicted boxes on the test set.Figures 21 and 22 show a sample of predicted versus ground-truth boxes and the predicted-box counts, respectively.Source: §4.4.3 · pp. 25–26, Figure 21 · p. 25, Figure 22 · p. 26
  • WFD-C had the highest AP score among the ensemble models, and its precision-recall curve is shown.The passage identifies WFD-C as the ensemble model with the highest AP score; Figure 23 presents its precision-recall curve.Source: §4.4.3 · p. 26, Figure 23 · p. 26
  • 0.8639 AP50 for WFD-C versus 0.754 for PAA on the Gazi University Hospital test set, a reported increase of 0.1099; WFD-C also had better AP results than YOLO across different test-set sizes.When the test amount was doubled, WFD-C achieved better AP results than PAA.Source: Table 11 · p. 27, §4.4.3 · p. 27

5. Conclusions and Future Work

The study reports WFD-C as its strongest wrist-fracture detector and highlights contributions to augmentation, evaluation, and ensemble design. It also identifies constraints on fracture coverage and model/input scope while outlining possible clinical applications.

5. Conclusions and Future Work

  • The WFD-C ensemble achieved the study’s highest average precision among 26 detection models, while the best weight-box-fusion ensemble improved AP by approximately 10%.The study does not give a baseline value for the approximate increase in this passage.Source: §5 · p. 27
  • The study-selected augmentation method is proposed as a reference for future work on similar medical X-ray images.Source: §5 · p. 27
  • The authors report calculating LRP parameters on medical data for the first time in the literature as part of their evaluation.Source: §5 · p. 27
  • The authors identify potential future hospital use and propose real-time support on portable devices, including for emergency settings and other frequently encountered bone fractures.Source: §5 · p. 28
  • Fractures of the other small wrist bones, including the scaphoid and lunate, were excluded from the study.Source: §5 · p. 28
  • Models unable to support 800 × 800 × 3 input images, including YOLO, were not used.Source: §5 · p. 28
  • Validation excluded the patient with fractures in both hands, and fracture labeling involved no more than three physicians with no disagreement reported.Source: §5 · p. 28
  • The dataset did not balance under-12 and adult patients, sex, right- and left-hand images, or fracture counts per image across train, validation, and test groups.Source: §5 · p. 28
Loading 2111.07355v3…