Computer Aided Diagnostic System for Tuberculosis Identification and Segmentation(Pixel-Level Detection) from Chest X-rays:
Tuberculosis (TB) is a contagious disease designated as foremost source of death across the world. Although TB can be treated with high change of cure, a large number of patients decrease for not getting timely and correct diagnosis, mainly due to insufficient radiography and radiologists. If left misdiagnosed and untreated, every active TB (sputum positive) case can contaminate 10 to 15 individuals in a single year. In 2016, about 10.4 million people infected with TB, resulted in 1.3 million deaths. Pakistan is 5th largest high-burden country in TB where about 510 million cases are reported yearly. Dealing with the TB challenge demands a diagnostic solution that is inexpensive, fast, precise, and easy to implement in undeveloped TB dominant areas. Currently the high burden TB countries are mostly relying on GeneXpertMTB-RIF (GXP) and sputum smear microscopy tests for TB diagnostics. Nevertheless, the expenses of a GXP test and lower efficacy of the smear test have led the World Health Organization (WHO) to suggest the use of chest radiography or Chest X-Rays CXRs for systematic diagnosis of TB. X-ray scans highlight irregularities in the lungs that can lead to diagnose TB. As digital X-ray has low operational expenses and rapid outcomes, it makes TB diagnostic affordable and quick. However, the correct diagnosis of TB from X-rays scans involves native expertise of radiology which is often defficient in TB occupied regions. Therefore, designing a CAD system for screening TB can be very beneficial for early TB diagnosis and leads to deterrence of demises from TB.
-
Classification of Tuberculosis from Chest-X-rays
-
Classification using Transfer-Learning with Pretrained-VGG16
- Implementation
- Statistical Results
- Visualization
-
Classification using Transfer-Learning with Pretrained-MobileNet-V2
- Implementation
- Statistical Results
- Visualization
-
Segmentation of Tuberculosis Tuberculosis Infectious Regions
- Dataset Description
- Segmentation with FCN-8 Architecture
- Implementation
- Visualization
These datasets contain de-identified chest X-Rays (CXRs) of normal and anomalous cases. The study is conducted on two datasets including Shenzhen and Montgomery County (MC) and is publicaly available Here. The Shenzhen dataset comprises of 662 CXRs where 326 are normal and 336 are anomalous (i.e. TB) CXRs. The dataset was collected from People's Hospital, Guangdong Medical College, Shenzhen, China. The MC dataset consists of 138 CXRs including 80 normal and 58 anomalous CXRs. Using 80:20 train/validation split I have split the total of 800 CXRs to 600 training set and 200 test/validation set. This splitted dataset can be downloaded at Here. I have also write a script to further make the Numpy-Files for the data. These Numpy-Files can be downloaded at Here. As the number of training instances are too less to train a good generalized model, That's why I have used some data-augmentation techniques with ImageDataGenerator.
 |
 |
|---|---|
| Confusion Matrix | Receiver Operating Characteristic (ROC) curve |
 |
 |
|---|---|
| Confusion Matrix | Receiver Operating Characteristic (ROC) curve |
To the best of my knowledge, there is no such work done for pixel-level detection of TB lesion in a CXR. In most of the previous works the authors tried to detect pixel level labels with weak supervision. One of the main reason for this is that there is no availability of pixel-level datasets for TB infectious regions and to acquire these is so time consuimg and costly process. We have also proposed a novel approach for the detection of pixel-level labels with Generative Adversarial Networks using weak-supervision. This work is available at my GitHub Repository.
Pixel-Level masks/labels for Tuberculosis Infectious Regions aren't publicaly available yet. A paper was arxived in 2016 in which the authors have hired radiologists to mask/label the infectious region of TB lesions in a chest-Xray. The authors of this paper have used these masks/labels to validate their proposed methodology of detecting the TB lesions with weak supervision using deconvolutional-feature stacking. I have somehow managed to get these masks/labels by requesting the authors of that paper. If you are a researcher or healthcare worker and you would like to get these masks/labels to validate your weakly supervised model or to propose a novel supervised model for pixel-level detection of TB lesion in a CXR, please reach out to zshnnisar@gmail.com. The pixel-labels are available only for Shenzhen Chest-Xrays. The original images and masks/labels were of size (500 * 500) but I have write a script to pad on edges for both images and masks/labels to get a size of (512 * 512).
