Deep learning models refer to a class of computing machines that can learn a hierarchy of features by building high-level attributes from low-level ones [1,2] , thereby automating the process of feature construction. One of these models is the well-known convolutional neural network (CNN) [2]. Consisted of a set of algorithms in machine learning, CNN comprises several (deep) layers of processing involving learnable operators (both linear and non-linear), and hence has the ability to learn and build high-level information from low-level features in an automatic fashion [3].
Stemming from biological vision processes, a CNN applies a feed-forward artificial neural network to simulate variations of multilayer perceptrons whereby the individual neurons are tiled in such a way that they respond to overlapping regions in the visual field [4]. As a direct result, these networks are widely applied to image and video recognition. Specifically, CNNs have demonstrated as an effective class of models for understanding image content, proffering state of the art results on image recognition, segmentation, detection and retrieval. For example, when trained with appropriate regularization, CNNs can achieve superior performance on visual object recognition tasks without relying on any hand-crafted features, e.g. SIFT, SURF. In addition, CNNs have been shown to be relatively insensitive to certain variations on the inputs [5]. Significantly, recent advances of computer hardware technology (e.g., Graphics Processing Unit (GPU)) have propitiated the implementation of CNNs in representing images.
Theoretically, CNN can be expressed in the following formulas. For example, for a set of training data , where image is in three-dimension (inclusive of RGB channel as the 3 rd dimension) and the indicator vector of affiliated class of the feature maps of an image, namely, , will be learnt based on CNN by solving Eq. ( 1).
(
Where refers to a suitable loss function (e.g. the hinge or log loss) and the selected classifier.
To obtain these feature maps computationally, in a 2D CNN, convolution is conducted at convolutional layers to extract features from local neighbourhood on the feature maps acquired in the previous layer. Then an additive bias is applied and the result is passed through a sigmoid function as formulated in Eq. ( 2) mathematically in order to obtain a newly calculated feature value at position on the feature map in the layer.
(
where the notations of those parameters in Eq. ( 2) are explained in Table 1.
hyperbolic tangent function index over the set of feature maps in the layer bias for the feature map in Eq. ( 1). value at the position (p, q) of the kernel connected to the kth feature map 2D position of a kernel
As a result, CNN architecture can be constructed by stacking multiple layers of convolution and subsampling in an alternating fashion. The parameters of CNN, such as the bias and the kernel weight are trained using unsupervised approaches [6,7].
This work is to response to the challenge task of automatic detection of Tuberculosis (TB) types using datasets that is organised by ImageCLEF as put forward in [8,9], part of CLEF conference to take place in Dublin [10,11].
As explained in [8], data are collected for the evaluation of ImageCLEF tuberculosis competition with 500 datasets of 3D CT lung images (512x512xdepth) for training and further 300 for testing. Among 500 datasets there are 140, 120, 100, 80 and 60 respectively for five TB types of for Infiltrative (type 1), Focal (type 2), Tuberculoma (type 3), Miliary (type 4) and Fibro-cavernous (type 5).
For the training data, each 3D dataset firstly undergoes pre-processing stage, whereby those artefacts are removed, i.e., slices that contain little visual content will be excluded. In this way, each dataset includes slices between 100 and 170. Then upon each slice, patches of varying sizes are created based on the lung boundary is created from its mask file [12]. Figure 1 illustrates the pre-processing by depicting the montage of original dataset (top-left), segmented (top-right) and patches segmented from one slice (bottom) that contains at least 80% of lung contents, checked by its mask. to be categorised, class 6 is created to include those patches that appear to be common among the first 5 classes, which again contains about 1000 patches, which is illustrated in Table 2. Figure 2 exemplifies the five types of patches that are applied in the training, which are all normalised into 30x30 pixels. As discussed above, type 6 remains an extra class to contain those common features shared between types 1 to 5 with patches randomly selected from type 1 to 3, and 5 without type 4, this is because the outstanding features in type 4 with miliary TB spread nearly every slice. Since the patches in each class are randomly generated by computers and selected manually, it is more likely that many patches belong to both of the five classes and class 6, which should not cause too much concerns as the final decision making is based upon the probabilities obtained for the first 5 classes. In addition, the patches belong to each subject will stay together when dividing between training and test data.
Table 2. The lists the number of patches in the training process.
The CNN network that has been designed for this task is built upon matConvNet package written using Matlab software1 . Seven layers of CNN have been designed
with input data of 30x30 pixels. The filter sizes for each layer are of (6,6), (4,4), (
) respectively. At layer 6, instead of scoring features into one of the six classes using Softmax approach that applies cross-entropy loss interpreting the scores as (unnormalised) log probabilities for each class, this study applies support vector machines (SVM) that adapts hinge loss to encourage the correct class to have a score higher by a margin than the other class scores. In this way, each class has a distinguishing boundary. For example, if a feature belongs to one of the two classes with probabilities of 0.51 and 0.49 respectively, the classification is not quite convincing. However, a score with a higher margin, e.g., 0.65, is more acceptable.
Within the existing of 500 datasets that have known ground truth, with reference of patch-wise, the classification results are given in Table 3 in percentage with an overall classification rate of 96%.
While the overall classification appears to be reasonable, with 86.49% accuracy rate, the result for Type 1 only sustains 50%. Since the testing datasets only include 8 subject samples in this type, the remaining work is to evaluate the results from real testing datasets when the ground truth is obtained. In order to obtain higher accuracy, it is recommended that medical knowledge should be embedded. Additionally, 3D segments should be also included to further enhance the characteristics that 3D datasets entail.