In this paper, we describe our model designed for automatic detection of diseases based on multimedia data collected in hospitals. Specifically, a two-stage learning strategy is designed to predict the diseases. In the first stage, a dimensionality reduction method called bidirectional marginal Fisher analysis (BMFA) is proposed to project the original data to the low-dimensional space, with the key discriminant information being well preserved. In the second stage, the multi-class support vector machine (SVM) is utilized on the low-dimensional space for detection. Experimental results demonstrate the eficiency of designed model.
We have evidenced an increasing trend of applying
multimedia processing and analysis methods, such as computer
vision and medical image processing, to assisting diagnosis,
detection and interpretation of medical abnormalities [
The 2017 MediaEval Medico Task aims to improve
detection and location of abnormalities through designing an
integrated approach combining information from both video
and image information, as well as other sensory information
and assistance of medical experts. The dataset used in this
task consists of 8,000 GI tract images that are annotated
and verified by medical doctors for the ground truth. More
details about the task requirements and the dataset can be
found in [
In order to perform eficient detection of diseases, we design a two-stage learning strategy. In the first stage, a manifold learning method called bidirectional marginal Fisher analysis (BMFA) is proposed to learn the compact representation of the original data, with the key discriminant information being well preserved. In the second stage, the multi-class support vector machine (SVM) is applied to the compact representation of each data point for detection. 2.1
Given the data matrix X = [x1, x2, ..., x], where denotes the number of data points in the training set, x ∈ R the high-dimensional feature vector of the -th data point, ∈ {1, 2, ..., 8} the set of possible labels, and l = [1, 2, ..., ] the set of true labels corresponding to each data point x. The proposed bidirectional marginal fisher analysis (BMFA) aims to learn the subspace of original features, in which the marginal discriminant information is well preserved.
A manifold learning algorithm typically involves the
definition of a neighborhood graph. The efectiveness of the
algorithm very often depends on how well the defined
neighborhood graph can preserve the local topological properties of
the data manifold. To achieve a robust neighborhood graph,
we use the following two-way connection criterion to
construct the graph [
The within-class adjacency matrix and between-class adjacency matrix are then defined as follows: A = A = {︃ {︃ exp(− ||x− 2x||2 ), if ∈ and ∈ and = , 0, otherwise exp(− ||x− 2x||2 ), if ∈ and ∈ and ̸= , 0, otherwise (2) where denotes the index set of the nearest neighbors of x, and is empirically set by = ∑︀=1 ||x − x ||2 / where x is the th nearest neighbor of x.
BMFA aims to learn a transformation matrix W ∈ R× where the following objective function is maximized: ∑︀,=1 A||W x − W = arg max ∑︀,=1 A||W x −
W
W x||22 .
W x||
The above optimization problem can be relaxed to the following ratio trace problem:
W = arg max (
W
where S = XLX and S = XLX , in which L = D −
A and L = D − A are the Laplacian matrices [
=1 A ( = 1, ..., ). The optimization problem in Eq. (4) could be solved by generalized eigen-decomposition.
For each high-dimensional data point x, we can obtain its low-dimensional representation by y = W x. 2.2
To predict diseases, we apply the multi-class support vector
machine (SVM) to the low-dimensional representation of
each data point. Specifically, we employ the one-against-all
strategy for the multi-class classification and use the -SVM
[
In this section, we report the experimental settings and the
evaluation results. The Kvasir Dataset is used to evaluate
the performance of the proposed model [
In the following step, we use the LIBSVM toolbox for
classification [
Table 1 shows the oficial evaluation results (provided by the task organizers) of the proposed model on the Kvasir dataset. From the table, we can see that the proposed model performs well, given the task is a multi-class classification with only a portion of the available data being utilized for training. This indicates that the learned subspace is able to capture the discriminant information of the dataset. Moreover, by reducing the original data to a very low-dimensional subspace, the learning process becomes more eficient.
In addition to the overall performance, we analyze the contribution of each dimension in the original feature space. The contribution of the -th dimension is defined as = ∑︀ ||, where denotes the -th eigenvalue, denotes the (, )-th element of W, and | · | denotes the absolute value operator. From Figure 1 we can see that the edge histogram and auto color correlogram features contribute more than the others, which indicates their importance in medical image classification. 4
This paper described the model designed for disease detection based on multimedia data. A novel dimensionality reduction algorithm dubbed bidirectional marginal Fisher analysis (BMFA) was presented to extract the discriminant information from the original feature space. After that, SVM was applied to the low-dimensional subspace for multi-class classification. Experimental results on the Kvasir dataset validated the efectiveness and eficiency of the proposed model.
This work was supported in part by the National Natural Science Foundation of China under Grant 61503317, and in part by the Faculty Research Grant of Hong Kong Baptist University (HKBU) under Project FRG2/16-17/032.