Computer Aided Diagnosis (CAD) of diseases has undergone large developments with the application of deep learning algorithms to detect the presence of diseases. This paper presents an approach for predicting the presence of tuberculosis, caverns and pleurisy in a set of 3D CT scans of the chests of patients, which is the key task of the ImageCLEF 2020 Tuberculosis challenge. We used the masks provided by the ImageCLEF organizers to segment the 3D CT images, made 2D projections of the segmented 3D images, and augmented them in order to balance the images in the dataset. An AlexNet based model is used to predict the probability of the presence of tuberculosis, caverns and pleurisy from these 2D projections. We achieved the eighth place out of all the teams who made a submission in this task, achieving a mean Area Under the Curve (AUC) score of 0.601 and a minimum AUC score of 0.432. An analysis of the results obtained by the authors following this approach presented, exploring the role of the model's complexity in reduction of the desired performance.
Tuberculosis
Computer Tomography
Tuberculosis (TB) is caused by bacteria (Mycobacterium tuberculosis) that most
often a ect the lungs of human beings [
This approach of analyzing 3D CT images of lungs for Tuberculosis consists of projecting the 3D image into 2D on the three planes - XY, YZ and XZ respectively. Following this, an AlexNet based model is used to predict the probability of a particular lung a ected by TB, the probability of the presence of Caverns, and the probability of the presence of Pleurisy, respectively. 2
The main task for ImageCLEF 2020 tuberculosis [
The working notes of the participating teams are to be published in the Proceedings of the 11th International Conference of the CLEF Association (CLEF 2020) [11]
The provided dataset, ImageCLEFmed Tuberculosis 2020, comprises of chest Computed Tomography (CT) scans of 403 TB patients. Out of these samples, 283 are designated for training and remaining 120 for testing. The CT images are digitized as a set of 2D slices and the distance between each of these 2D slices can vary between 0.5 - 5mm in any axis, depending upon the resolution of the 3D image. In our case, these slices are stored in the compressed NIfTI format. 3 3.1
The data preprocessing task is performed similar to the method used by
Vitali Liauchuk [
Using random selection on generated images, we take a ratio of 3:1 for training and validation sets, respectively. We observed that the data is skewed, especially with pleurisy where the number of una ected samples outweigh the a ected ones. We augment the images using random ipping and rotation to balance the dataset wherever possible. 3.2
The 2D projections of the lung images does the heavy-lifting of extracting the important features required to mark the lesion or deformity in the lung. These features can easily be detected by a simple neural network. It would thus reduce the problem to detection of these lesions in the 2D projections. After prepossessing and data augmentation, we observed that recent state-of-the-art image classi cation neural networks over t quickly because of their inherent complex feature extraction. Hence we choose a relatively simple AlexNet-based solution to train for prediction.
AlexNet [14] is a CNN with stacked convolutional layers. It consists of 11 11, 5 5, 3 3, convolutions, max pooling, dropout, data augmentation, ReLU activations and SGD with momentum. It attaches ReLU activations after every convolutional and fully-connected layer.
Keeping the underlying AlexNet structure, we changed the number of out channels for ve ConvBlocks to 64, 128, 256, 512, 512 respectively, each one of the ConvBlock is then followed by ReLU and MaxPooling of stride 2.
The basic block of our architecture, ConvBlock, is a simple 2D convolutional layer, followed be a ReLU layer and a MaxPool2D layer. The simple 2D convolutional layer has a kernel size of 2 and stride of 2, and the MaxPool2D layer has a kernel size of 3 and stride of 1. We added Dropout, linear, ReLU and again Linear layer after the ve ConvBlocks, in order to build a suitable model for Binary classi cation, which is then changed into probability distribution.
The architecture of the proposed model used for this task is shown in Fig 1. The input to this model is a pseudo-RGB image which consists of three channels. The output is the probability distribution of the target variable that signi es the probability of lung being a ected. We use PyTorch [16] to construct and train the model. Starting from left, the number in parentheses indicates the number of output channels for each of the ConvBlocks. Following ConvBlocks is the dropout layer, which has an activation probability of 0.5. The following Linear layer is a fully connected layer, the number of nodes present in whom are represented by the bracket enclosed numbers. To introduce non linearity among the two Linear layers, we use ReLU activation function. SoftMax layer follows the nal Linear layer, which converts the output of the model into a probability distribution. We trained separate models for Lung, Caverns and Pleurisy respectively. We observed that the patterns associated with the disease do not change with respect to the side of the lung, whether left or right. Therefore, we use left and right a ected lung samples from the dataset collectively, to train one model. Separate models were trained for all 3 di erent 2D projections, thus for each subtask say, LungA ected, we train one model for XY-projection, one for YZ-projection and one for ZX-projection. We approach the training as a classi cation task. We used Kaiming initialization [17] to initialize the weights of the network and trained the model using a mini-batch size of 4. We used Categorical cross entropy loss and a learning rate of 10-5 while training. We used Adam optimizer with weight decay of 0.0005. The loss and accuracy saturated after 5 to 7 epochs for all the training sets. In the end, we have a total of 9 trained models, XY, YZ, ZX - projection model for each of Lung, Caverns and Pleurisy detection subtask.
We performed test time augmentation while predicting the probability values. The XY, YZ and ZX projections were taken and given as input to the model separately. The trained model does not di erentiate in left and right lung, so we pass both projections one after the other to the same model and store the prediction values separately. The resulting outputs were taken and given as input to a SoftMax layer, which converts them into a probability distribution. This provides us with the probabilities for the predictions. While predicting the values individually for XY, YZ and ZX projections, the prediction scores for LungAffected and Caverns were almost similar, but the scores of Pleurisy values for the XY and YZ axes were low while the ZX-axis score was substantially higher. Hence, we used only the ZX-projection for Pleurisy and mean of all the projections for others.
We did two submissions for the task. The primary di erence between the two models was the number of channels and inclusion of test time augmentation, which gave a signi cant improvement to the scores, bringing the mean AUC to 0.601 shown in Table 1. We mentioned the approach and methodology in section 3. With our best prediction, we achieved the eighth place out of the nine teams who made a submission and the results are shown in Table 2. Our best submission, Run 2 (JBTTM) achieved a mean AUC score of 0.601 and min AUC score of 0.432, Run 1 achieved better min AUC score of 0.471 but poor Mean AUC score of 0.484. The top ranking team SenticLab.UAIC scored 0.924 and 0.885 on Mean and Min AUC respectively.
The method of using AlexNet on 2D projections suggested promising results. We detected TB lesions with a substantial certainty and performance in case of caverns is good. However, it did not perform as expected in detecting pleurisy, partly because it is a non-localized phenomenon like a TB lesion. We can also attribute the reason for a lower score to the large number of channels in the network which added unnecessary complexity to the model and caused it to over t that resulted in poor performance of our model, especially in detecting pleurisy.
In this work, we experimented with an AlexNet Based Model to predict the probability of a lung a ected by TB, the probability of the presence of Caverns, and the probability of the presence of Pleurisy respectively. Data preprocessing and augmentations are done as described in the previous sections to prepare the images for the deep neural network. The performance of the models submitted are measured using mean AUC and minimum AUC. Our second submission run has achieved 0.602 and 0.432 for the speci ed measures. The performance when compared with the other submissions of this task has shown reasonable yet improvable results. Our team, JBTTM, has achieved eighth place out of the nine teams who have submitted their runs. We observed that the increased model complexity led to over tting and thereby pulling down the model performance. A smaller and simpler model, along with proper regularization techniques, could be used in order to achieve a better result.