Recently, deep Convolutional neural networks (CNN) have obtained significant results in volumetric medical image classification task. However, a number of studies have claimed that replacing the trainable classifier (the conventional softmax function) of a CNN model with an SVM classifier can improve the classification performance. In this paper, we investigate the use of a hybrid approach which combines 3D CNN and a linear SVM to tackle the task of Multi-Drug Resistant detection.
Tuberculosis (TB) is a serious infectious disease usually caused by the
Mycobacterium tuberculosis (MTB) [
The motivation of ImageCLEFtuberculosis 2018 task [
Recently, deep Convolutional neural networks (CNNs) have obtained significant
results in volumetric medical image classification task [
The remainder of this paper is organized as follows: the dataset is described in Section 2, the proposed methodology is described in Section 3, results are presented and discussed in Section 4, and finally Section 5 draws some conclusions and future work.
• MDR detection dataset: MDR detection is a Binary classification task, where the goal is to predict the probability for each of the two classes: Drug Sensitive (DS) and Multi-Drug Resistant (MDR). As shown in the table below a dataset of 3D CT images is used with 259 of training examples, and 236 of test cases. This section elaborates on the details our proposed approach to address the MDR detection task from ImageCLEFtuberculosis 2018. As shown in the fig.1, our system is designed to (1) automatically extract prominent features from 3D CT scans using a 3D CNN, and then (2) classify them into predefined categories with the Linear SVM classifier. The input 3D CT scans are stored in Niftii file format, where the number of slices is around 50- 400, and the image size per slice of 512x512 pixels. However, slices variations can be problematic for automatic analysis using CNN. For that we have standardized the size of input images to 100x100x50.
3D CNN as features Extractor: Generally, CNN architecture consist of three main types of layers (convolutional layers, pooling layers and fully connected layers) that are stacked on top of each other to form a full CNN model. 1) Convolutional layer: the convolutional layer is always the first layer of CNN, it serves as features extractor, and it consists of several features maps. Each neuron in a feature map is connected to a small region, called the local receptive field, through a set of shared weights and a single shared bias.
Two main advantages of using Convolutional layer instead of fully connected
layer. The first one is parameters sharing; where sharing weights and bias are
equal for all neurons in a feature map, which greatly reduces the number of
parameters that in turn will result in faster training, and, ultimately, will help to build
deep networks [
CNNs are mainly developed for 2D images, however in the medical domain most
images are 3D tensors. The traditional way to solve this problem is to convert 3D
image into 2D representation, but this method leads to a loss of information, which in
turn limits the performance of the overall system [
Name Data Conv3D_1 Max pool3D_1 Conv3D_2 Max Pool3D_1 FC1 FC2 As shown in the table above, our 3D CNN has about 130 million parameters. It consists of two convolutional layers with the same padding and 64 filters with a kernel size of 5x5x5; each Convolutional layer is followed by a max-pooling layer with overlapping 3 × 3 × 3 windows, and two fully connected layers with 1024, and 2 neurons respectively.
At the first fully connected layer and in each convolutional layer, we have chosen
to use a Rectified Linear Unit (ReLU)[
Our 3D CNN was trained on the training set, with “Categorical Cross-entropy” as
loss function and Adam [
However, this trained model is used only as features extractor. Therefore, we have replaced the fully connected layers (FC1 and FC2) with the SVM classifier, which takes the flattened output from the rest layers as a new feature vector.
Support Vector Machine (SVM) classifier:
The SVM [
To predict the probability of a TB patient having resistant form of tuberculosis, we have used the SVM classifier with a linear kernel. This classifier takes as input the extracted features from the pre-trained 3D CNN (with a size of 127050), and trained on the whole training set.
In this section, we present the performance results of our model on the test set:
Multi-Drug Resistant (MDR) Detection:
MDR detection task consists on assigning the probability of each patient having
MDR. For that, two metrics was used to evaluate model performance: Accuracy and
Area Under the Curve (AUC) obtained from the ROC-curves produced with the
submitted probabilities. The following table provides the results obtained by our model
and the best run for the task of MDR detection.
Our model gives good results on MDR detection task with an AUC of 0.5029 and an
accuracy of 0.4576. However, the best model has achieved an accuracy of 0.6144 and
AUC of 0.6178. For that we aim to improve the performance of our model by using
the provided masks[
In this paper, we investigate the use of a combination of 3D-CNN and SVM classifier to tackle the problem of MDR detection. Where, the 3D-CNN is used to extract relevant features from patients CT scans, while the SVM is used for the classification step. However, the test results of our submitted run are not satisfactory comparatively with the results obtained by the best model. Therefore, we aim to improve the performance of our model by adding more training data using the provided masks to reduce the size of 3D volumes before resampling them. With this we can focus on the lung tissue and we can reduce the noise from the rest of the body.