Analyzing chest X-rays is a common clinical approach for diagnosing pulmonary and heart diseases. However, foreign objects are occasionally presented on chest X-ray images, especially in rural and remote locations where standard filming guidances are not strictly followed. Foreign objects on chest X-rays may obscure pathological finds, thus increasing false negative diagnosis. They may also confuse junior radiologists from real pathological findings, e.g. buttons are visually similar to nodules on chest X-ray, thus increasing false positive diagnosis. Based on statistics from our (JF Healthcare) telemedicine platform, nearly one third of the uploaded chest X-rays from township hospitals do not qualify for diagnosis, and interference from foreign objects is the largest contributing factor. Before we develop our own detection system, our radiologists have to manually check the quality of the chest X-ray and promote re-filming if foreign objects within the lung field are presented. An algorithm based foreign object detection system could automatically promote re-filming, thus significantly reduce the cost and save radiologists’ time for more diagnosis, which has been noticed before (L Hogeweg, 2013, Medical Physics).
We provide a large dataset of chest X-rays with strong annotations of foreign objects, and the competition for automatic detection of foreign objects. Specifically, 5000 frontal chest X-ray images with foreign objects presented and 5000 frontal chest X-ray images without foreign objects are provided. All the chest X-ray images were filmed in township hospitals in China and collected through our telemedicine platform. Foreign objects within the lung field of each chest X-ray are annotated with bounding boxes, ellipses or masks depending on the shape of the objects.
Detecting foreign objects is particularly challenging for deep learning (DL) based systems, as specific types of objects presented in the test set may be rarely or never seen in the training set, thus posing a few-shot/zero-shot learning problem. We hope this open dataset and challenge could both help the development of automatic foreign objects detection system, and promote the general research of object detection on chest X-rays, as large scale chest X-ray datasets with strong annotations are limited to the best of our knowledge.
object-CXR challenge is going to be hosted at MIDL 2020.
5000 frontal chest X-ray images with foreign objects presented and 5000 frontal chest X-ray images without foreign objects were filmed and collected from about 300 township hosiptials in China. 12 medically-trained radiologists with 1 to 3 years of experience annotated all the images. Each annotator manually annotates the potential foreign objects on a given chest X-ray presented within the lung field. Foreign objects were annotated with bounding boxes, bounding ellipses or masks depending on the shape of the objects. Support devices were excluded from annotation. A typical frontal chest X-ray with foreign objects annotated looks like this: We randomly split the 10000 images into training, validation and test dataset:
training 4000 chest X-rays with foreign objects presented; 4000 chest X-rays without foreign objects.
validation 500 chest X-rays with foreign objects presented; 500 chest X-rays without foreign objects.
test 500 chest X-rays with foreign objects presented; 500 chest X-rays without foreign objects.
We provide object-level annotations for each image, which indicate the rough location of each foreign object using a closed shape.
Annotations are provided in csv files and a csv example is shown below.
image_path,annotation /path/#####.jpg,ANNO_TYPE_IDX x1 y1 x2 y2;ANNO_TYPE_IDX x1 y1 x2 y2 ... xn yn;... /path/#####.jpg, /path/#####.jpg,ANNO_TYPE_IDX x1 y1 x2 y2 ...
Three type of shapes are used namely rectangle, ellipse and polygon. We use
For rectangle and ellipse annotations, we provide the bounding box (upper left and lower right) coordinates in the format
x1 y1 x2 y2where
For polygon annotations, we provide a sequence of coordinates in the format
x1 y1 x2 y2 ... xn yn.
Our annotations use a Cartesian pixel coordinate system, with the origin (0,0) in the upper left corner. The x coordinate extends from left to right; the y coordinate extends downward.
The training and validation dataset can be accessed here at Google Drive. The imaging data has been anonymousized and free to download for scientific research and non-commercial usage.
We use two metrics to evaluate the classification and localization performance of foreign objects detection on chest X-rays: Area Under Curve (AUC) and Free-response Receiver Operating Characteristic (FROC).
For the test dataset, each algorithm is required to generate a
prediction_classification.csv file in the format below:
image_path,prediction /path/#####.jpg,0.90 /path/#####.jpg,0.85 /path/#####.jpg,0.15 ...
where each line corresponds to the prediciton result of one image. The first column is the image path, the second column is the predicted probability, ranging from 0 to 1, indicating whether this image has foregin objects or not.
We use AUC to evaluate the algorithm performance of classifying whether each given chest X-ray has foreign objects presented or not. AUC is commonly used to evaluate binary classification in medical imaging challenges. For example, the CheXpert competition for chest X-ray classification, and the classification task within CAMELYON16 challenge for whole slide imaging classification. We believe AUC is adequate enough to measure the performance of the classification task of our challenge, especially given our positive and negative data is balanced.
For the test dataset, each algorithm is required to generate a
prediction_localization.csv file in the format below:
image_path,prediction /path/#####.jpg,0.90 1000 500;0.80 200 400 /path/#####.jpg, /path/#####.jpg,0.75 300 600;0.50 400 200;0.15 1000 200 ...
where each line corresponds to the prediciton result of one image. The first column is the image path, the second column
is space seperated 3-element tuple of predicted foreign object coordinates with its probability in the format of (probability x y), where x and y are the width and height coordinates of the predicted foreign object. It is allowed to have zero predicted 3-element tuple for certain images, if there are no foreign objects presented. But please note the
, after the first column even if the prediction is empty.
We use FROC to evaluate the algorithm performance of localizing foreign obects on each given chest X-ray. Because our object annotations are provided in different format, i.e. boxes, ellipses or masks depending on radiologists’ annotation habits, it’s not suitable to use other common metric, such as mean average precision (mAP) in natural object detection with pure bounding box annotations. FROC is more suitable in this case, since only localization coordinates are required for evaluation. FROC has been used as the metric for measuring medical imaging localization, for example, the lesion localization task within CAMELYON16 challenge, and tuberculosis localization in (EJ Hwang, 2019, Clinical Infectious Disease).
FROC is computed as follow. A foregin object is counted as detected as long as one predicted cooridinate lies within its annotation. The sensitivity is the number of detected foreign objects dividide by the number of total foreign objects. A predicted coordinate is false positive, if it lies outside any foreign object annotation. When the numbers of false positive coordinates per image are 0.125, 0.25, 0.5, 1, 2, 4, 8, FROC is the average sensitivty of these different versions of predictions.
froc.py provides the details of how FROC is computed.
We provide a baseline result in this Jupyter Notebook.
We host the online submission system at codalab
JF Healthcare is the primary organizer of this challenge