Tuberculosis (TB) is an air-borne disease, which affects the lungs and often spreads through sputum. According to the report of World Health Organization 9 million people world-wide are affected with TB. Tuberculosis can be cured easily when diagnosed in its early stage and with accurate CT Analysis. As an effort to form a technical forum for effective analysis and diagnosis, ImageCLEF released the Tuberculosis 2019 tasks, each dealing with one aspect of understanding and tackling the disease. We have taken up one sub-task that aims at assessing the severity of the tuberculosis disease as low or high. The task is implemented using a deep neural network approach using 2-D Convolutional Neural Network (CNN) with appropriate preprocessing. The CT volumes are segmented with the provided masks and further pre-processed with the aid of med2image, a python utility to obtain slices of CT scans, prior to training the model. The best run of the proposed CNN model resulted with an accuracy of 0.607 and an AUC of 0.626. The achieved result is placed 9th in the overall leaderboard of the ImageCLEF 2019 Tuberculosis challenge for severity scoring .
Tuberculosis (TB) is an airborne disease that affects the lungs. Often spread through
sputum, cough and infected droplets, it is quite widespread affecting about 9 million
world-wide. The treatment depends upon the degree of infection, i.e the severity [
The sections span across following: Section 1 gives a brief introduction about the importance of this problem and the necessity to find the severity of tuberculosis. Section 2 gives a glimpse of the dataset and how it is spread across the two classes and Section 2.1 details about the data preprocessing procedures. Section 3 explains the proposed model using convolutional neural network with the parameters chosen for analysis. In Section 4, the results of various runs are discussed. Finally, Section 5 concludes this paperwork and looks into the futuristic aspects for further improvisation of the proposed model.
In this edition of ImageCLEF 2019 TB tasks, the dataset contains 335 chest CT
scans of TB patients along with a set of clinically relevant metadata, where data of
218 patients are used for training and 117 for testing. For all patients, 3D CT images
are stored in the compressed NIFTI (Neuroimaging Informatics Technology
Initiative) file format with a slice size of 512×512 pixels and the number of slices varies
from 50 to 400 for each patient. This file format stores raw voxel intensities in
Hounsfield Units (HU) as well the corresponding image metadata like image
dimensions, voxel size in physical units, slice thickness, etc. The selected metadata
includes the following binary measures: disability, relapse, symptoms of TB,
comorbidity, bacillary, drug resistance, higher education, ex-prisoner, alcoholic, smoking
and severity score ranges from 1 to 5 assigned by medical doctors. To treat this task
as a binary classification problem, the severity scores are grouped as high severity
with scores 1, 2 and 3, and low severity with scores 4 and 5. Moreover, for all
patients automatic extracted masks of the lungs are provided. In Table 1, the number of
patients of each severity class in the training set and the number of patients in test
set is given [
Convolutional Neural network takes an image as input, passes it through a series of
convolutional layers, nonlinear activation layers, pooling (downsampling) and a
fully connected layer to output the classification labels. It differs from normal neural
network in two aspects: atleast one convolutional layer and filters. The model for TB
severity scoring is created using 2-D convolutional neural network using software
libraries Keras [
The CNN model is trained by varying the hyperparameters such as number of filters, epochs and optimizers. The runs had a filter size of 64×64 with batch size 32 and loss type as binary cross entropy. The difference in the accuracy is brought by changing the epoch value and the optimizers such as Adam and RMSProp. The different runs of CNN model by varying the hyperparameters is given in Table2. binary cross entropy binary cross entropy binary cross entropy RMSProp
Adam
RMSProp
RMSProp Number of epochs Loss type
Optimizer
The intermediate visualization of convoluation layers and max pooling is shown in Figure 5, for Run1, Patient ID 181 and slice number 65.
The result of submitted four runs are listed in Table 3, for training, validation and test dataset with necessary parameters. In testing, 31 slices per patient is considered as similar to training and validation, for all 117 patients. The probability of high severity for each patient is calculated from the average of “probability of high” of all 31 slices of the specific patient. For example, the class probability of patient ID 77 in testing, for slice number 60 is represented as [0.1.]. Here, “0.” is the probability of having low severity and “1.” is the probability of having high severity. We find the probability of having high severity for each of 31 slices of the patient and then computed the average of it. The average probability of high severity for patient ID 77 in each run is given in Table 4.
The test dataset results are evaluated using two metrics namely accuracy and Area Under ROC Curve (AUC) and ranking is carried out among the participated teams.
For better visualization, the same information is plotted and shown as graphs in Figures 6 and 7 using Tableau Tool. In Figure 6, the value of evaluation metrics for test set is given for all four runs. From the graph, it is clearly visible that run 3 has higher AUC and accuracy than remaining runs. In Figure 6, the accuracy for training, validation and testing dataset are given for all four runs. From the graph, it is clearly visible that run 3 has higher value in all the cases. In addition, run 4 has higher training and validation accuracy, but the test accuracy is low than run 2, might have occurred due to the chosen filter size of each layer.
In this paper, analysis of severity scoring (SVR) subtask for lung tuberculosis using 2D Convolutional Neural Network is implemented. The classification results obtained for the given set 3D CT Images are submitted for evaluation. In our approach, preprocessing of the dataset has been carried out to convert the images into 2D slices, and the images are split into training and validation set. The proposed model is built using CNN, trained and validated using tuning the hyperparameters for four different runs. From the runs submitted, the primary run is ranked 9th place among the team participations.
CNN is a preferred approach, since it facilitates automatic detection of the low level and high level features, from large training dataset. However, a large dataset might prove disadvantageous in terms of memory during the training phase. This can be overcome by the use of a GPU and choosing optimal hyperparameters. In future, the proposed model can be improvised by considering all the slices of the CT images, to build the train model using GPU and transfer learning approach.