TBX11K: Tuberculosis X-ray dataset contains 11200 X-ray images with corresponding bounding box annotations for tuberculosis (TB) areas, while the existing largest public TB datasets have much fewer X-ray images with corresponding image-level annotations. All images are with a size of 512x512. There are five main categories in this dataset: healthy, sick_but_non-tb, active_tb, latent_tb, and uncertain_tb. Also, the dataset includes the active_tb&latent_tb categorie. The authors split this dataset into train, val, and test sets, consisting of 6600, 1800, and 2800(3302) X-ray images, respectively. The proposed dataset enables the training of sophisticated detectors for high-quality computer-aided tuberculosis diagnosis (CTD). Authors reform the existing object detectors to adapt them to simultaneous image classification and TB area detection. These reformed detectors are trained and evaluated on the proposed TBX11K dataset and serve as the baselines for future research.

As a serious infectious disease, tuberculosis is one of the major threats to human health worldwide, leading to millions of deaths every year. Although early diagnosis and treatment can greatly improve the chances of survival, it remains a major challenge, especially in developing countries. Computer-aided tuberculosis diagnosis is a promising choice for TB diagnosis due to the great success of deep learning.

active_tb&latent_tb refers to X-rays that contain active and latent TB simultaneously. active_tb and latent_tb refer to X-rays that only contain active TB or latent TB, respectively.

Labels for active_tb:

• ActiveTuberculosis
  latent_tb:
• ObsoletePulmonaryTuberculosis
  uncertain_tb refers to TB X-rays whose TB types cannot be recognized under today’s medical conditions. Uncertain TB X-rays are all put into the test set.

This is the distribution of the areas of TB bounding boxes. The left and right values of each bin define its corresponding area range, and the height of each bin denotes the number of TB bounding boxes with an area within this range. Note that X-rays are in the resolution of about 3000×3000. However, the original 3000×3000 images will lead to a storage size of over 100GB, which is too large to deliver. On the other hand, the authors found that the resolution of 512 × 512 is enough to train deep models for TB detection and classification. In addition, it is almost impossible to directly use the 3000 × 3000 X-ray images for TB detection due to the limited receptive fields of the existing CNNs. Therefore, the authors decided to only release the X-rays with a resolution of 512×512. For a fair comparison, they recommend all researchers use this resolution for their experiments.

Here are the answers from the dataset’s authors to questions about how the GT labels were created:

• Are latent TB cases biologically confirmed? (by IFNg testing or tuberculin skin testing)

Yes, they are. Both active and latent TB cases are biologically confirmed
using the hospitals’ accurate clinical diagnosis technology, of course,
in the image level.

• If the cases are biologically positive for active TB, but does has CXRs regions suspicious for latent TB only, are they labeled as latent TB or active TB?

As clarified in section 3.1.3, the annotation is conducted under a 
double-check rule: “Specifically, each TB X-ray is first labeled by a 
radiologist who has 5-10 years of experience in TB diagnosis. Then, his 
box annotations are further checked by another radiologist who has >10 
years of experience in TB diagnosis. They not only label bounding boxes 
for TB areas but also recognize the TB type (active or latent TB) for 
each box. The labeled TB types are double-checked to make sure that they 
are consistent with the image-level labels produced by the golden 
standard. If a mismatch happens, this X-ray will be put into the 
unlabeled data for re-annotation, and the annotators do not know which 
X-ray was labeled wrong before. If an X-ray is incorrectly labeled twice, 
we will tell the annotators the gold standard of this X-ray and ask them 
to discuss how to re-annotate it.” Therefore, the final TB type must be 
consistent with the golden standard.

• If the cases are biologically positive for active TB, but does not contain CXRs regions that are not suspicious for active TB, how are they labeled?

The answer is similar to that of the above (2). The image-level labels 
are confirmed using the hospitals’ accurate clinical diagnosis technology 
and thus reliable. In our annotation process, our experienced 
radiologists did not happen to the situation that you said, after 
discussion, the annotation still cannot be consistent with the gold 
standard.

• How can the cases be both active TB and latent TB?

Of course, one TB region is either active TB or latent TB. However, note 
that an X-ray would contain both active TB and latent TB regions. In our 
dataset, each TB box only has one label of being active or latent, but a 
TB X-ray would have both active TB and latent TB labels.

The proposed TBX11K dataset is much larger, better annotated, and more realistic than existing TB datasets, enabling the training of deep CNNs. First, unlike previous (Schenzen dataset and CXR digital image datasets) that only contain several tens/hundreds of X-ray images, TBX11K has 11,200 images that are about 17× larger than the existing largest dataset, i.e., Shenzhen dataset, so that TBX11K makes it possible to train very deep CNNs. Second, instead of only having image-level annotations as in previous datasets, TBX11K annotates TB areas using bounding boxes, so that future CTD methods can not only recognize the manifestations of TB but also detect the TB areas to help radiologists for the definitive diagnosis. Third, TBX11K includes four categories of healthy, active_tb, latent_tb, and sick_but_non-tb, rather than the binary classification for TB or not in previous datasets, so that future CTD systems can adapt to more complex real-world scenarios and provide people with more detailed disease analyses.

The ground truth of the test set will not be released, because it is an online competition for computer-aided tuberculosis diagnosis. To promote the development of this field, the authors suggest you use the train set for training and the val set for validation when tuning your model. When you submit your results of the testing set to the authors’ server, they suggest you train your model on the train+val set and test on the test set.