Deep Learning Models for Detecting Pulmonary Edema Features in Chest X-ray Radiographs
Description
Introduction: We present a collection of deep-learning models designed to accurately detect radiographic edema features. These models underwent rigorous training and testing using a dataset consisting of chest X-ray (CXR) images from the MIMIC database.
Dataset Description: A total of 1000 radiographs from 741 patients with features suggestive of edema were randomly selected for manual annotation by a qualified clinician with more than 10 years of radiology experience. The clinician meticulously labeled specific radiological features associated with pulmonary edema, such as cephalization, Kerley lines, peribronchial cuffing, pleural effusions, bat wings, and infiltrates. An example of the annotation methodology for different edema severity labels is shown in the image titled "Annotation methodology".
Model Selection: For the detection of radiological features, we carefully selected and trained the latest state-of-the-art object detection models, including TOOD, GFL, PAA, SABL, FSAF, Cascade RPN, ATSS, and Faster R-CNN (Table 1). Detailed information about these models is provided in Table 1. Model training and testing were performed on an Amazon Web Services P3 instance, specifically the p3.2xlarge instance, equipped with an 8-core Intel Xeon CPU E5-2686 v4 @ 2.30GHz, 61GB of RAM, and an Nvidia Tesla V100 GPU with 16GB of memory.
Model Comparison: We conducted a comprehensive evaluation of the studied object detection models to assess their performance in detecting radiological features, including cephalization, Kerley lines, effusion, infiltrate, and bat wings. These models underwent rigorous testing and their performance was quantified using two key metrics: latency and mAP (mean average precision). These metrics, presented in Table 2 and the accompanying image labeled "Model performance", provide valuable insight into the ability of these models to accurately and consistently identify specific radiological features.
Access to the Study: For more comprehensive information about this study, please visit our GitHub repository at https://github.com/ViacheslavDanilov/edema_quantification and our Zenodo dataset repository at https://zenodo.org/doi/10.5281/zenodo.8383776.
Table 1. Technical description of the pulmonary edema feature detection networks
Model | Release date | Backbone | GPU memory*, Gb | MAC, G | Parameters, M | Model size, Mb | mAP on COCO |
---|---|---|---|---|---|---|---|
TOOD | 2021 | ResNet-101 | 2.5 | 192.4 | 53.4 | 640 | 0.493 |
GFL | 2020 | ResNeXt-101 | 2.6 | 303.1 | 53.3 | 639 | 0.481 |
PAA | 2020 | ResNet-50 | 2.2 | 295.5 | 51.1 | 612 | 0.451 |
SABL | 2019 | ResNet-101 | 2.8 | 358.9 | 55.3 | 663 | 0.436 |
FSAF | 2019 | ResNeXt-101 | 3.8 | 461.6 | 93.9 | 1130 | 0.424 |
Cascade RPN** | 2019 | ResNet-50 | 1.9 | 232.2 | 41.9 | 501 | 0.404 |
ATSS | 2019 | ResNet-50 | 2.2 | 295.5 | 51.1 | 612 | 0.415 |
Faster R-CNN** | 2015 | ResNet-50 | 1.5 | 208.8 | 41.5 | 495 | 0.421 |
* The GPU memory allocated to the model during training with a batch size of 1.
** The exact numbers of MACs and parameters for Faster R-CNN and Cascade RPN are not explicitly stated in the existing literature. As a result, the numbers presented here are approximations and calculations based on available information.
Table 2. Metrics for the studied pulmonary edema feature detection networks
Model | Latency*, ms/image | APcephalization | APkerley | APeffusion | APinfiltrate | APbat | mAP |
---|---|---|---|---|---|---|---|
TOOD | 89 | 0.527 | 0.389 | 0.515 | 0.183 | 0.918 | 0.506 |
GFL | 95 | 0.524 | 0.386 | 0.574 | 0.274 | 0.913 | 0.534 |
PAA | 104 | 0.533 | 0.235 | 0.564 | 0.287 | 0.912 | 0.506 |
SABL | 62 | 0.526 | 0.395 | 0.599 | 0.395 | 0.926 | 0.568 |
FSAF | 85 | 0.509 | 0.377 | 0.488 | 0.286 | 0.892 | 0.510 |
Cascade RPN | 56 | 0.532 | 0.417 | 0.540 | 0.296 | 0.917 | 0.540 |
ATSS | 56 | 0.506 | 0.342 | 0.533 | 0.359 | 0.919 | 0.532 |
Faster R-CNN | 42 | 0.498 | 0.396 | 0.507 | 0.235 | 0.907 | 0.509 |
* Latency represents the average time required for a single image to be processed by a single model. When employing these single-class models in an ensemble, the number of models used will increase proportionally to the number of classes utilized in the ensemble. In our specific scenario, we employ five single-class models in the ensemble, resulting in a fivefold increase of processing time.
Files
Model performance.png
Files
(24.5 GB)
Name | Size | Download all |
---|---|---|
md5:e4cdb01892519a2b9bb5bfd54d65d049
|
9.2 MB | Preview Download |
md5:6649d16e6cc332ec5bebe3ec4acd22ff
|
2.9 GB | Preview Download |
md5:c65ea8028cae5c9b7945a822434cd309
|
2.3 GB | Preview Download |
md5:472b318ffeea15920540c1b6cb401bd1
|
2.3 GB | Preview Download |
md5:7486f6491e23e0beab13da12bad0b112
|
5.2 GB | Preview Download |
md5:3983c2a106f6dbdd41605d49f9af1e0c
|
3.0 GB | Preview Download |
md5:1de943c3907bf28e7c73c9e061cf5432
|
2.3 MB | Preview Download |
md5:6404179f65133bb821f70b9c69a58931
|
2.9 GB | Preview Download |
md5:1bcf0e53f4fe6351ea731a09c530af7b
|
3.0 GB | Preview Download |
md5:c82af120d94a653c9642a887a089c87a
|
2.9 GB | Preview Download |
Additional details
Identifiers
Software
- Repository URL
- https://github.com/ViacheslavDanilov/edema_quantification
- Programming language
- Python
- Development Status
- Active