Data Mining Techniques in Biomedical Imaging
April 7, 2016 | Blog, Uncategorized
Today in almost every engineering field that we name, biomedical, energy, chemical, electrical etc, applications of data mining techniques are gaining more significance. This blog discusses about the application of these techniques for diagnosing diabetic retinopathy. Diabetic retinopathy is a disease that occurs in diabetic patients due to damage in blood vessels. If not diagnosed at right time, it may lead to blindness. The diabetic retinopathy and its stages can be graded using the patient data in the form of fundus images. The retinal fundus images give details of the inner lining of the eye which includes the retinal pigment epithelium, the sensory retina, Bruch‘s membrane and the choroid. A data mining model to predict whether a patient is infected by this disease or not would help a physician to improve the diagnostic accuracy. Fig 1 shows sample normal, moderately infected and severely infected retinal fundus images respectively.
Fig 1.Sample Retinal Fundus Images
There are two phases in developing a predictive data mining model namely, training and testing. Training is done using labeled examples, when the learning algorithm identifies patterns in the training data to differentiate each category of examples. The testing phase validates the trained model using labeled examples. A rule of thumb is to use 70% of the data collected to train the model and 30% to validate it. If the predictive accuracy of the model is found acceptable, it can be used in future predictions.
To develop a predictive data mining model for diagnosing diabetic retinopathy, firstly a repository of retinal images of different categories like normal, moderately infected and severely infected patients are collected with the help of a retinal camera. Using domain experts who are ophthalmologists these images are labeled
as normal, moderate and severe. Any number of categories can be used to differentiate the severity level of the disease. Statistical features are then extracted from these labeled retinal images like mean, skewness, kurtosis, etc. These features are then used by a machine learning algorithm to differentiate each category of retinal images.
Apart from engineering these techniques are used by business organizations to predict new business strategies for increasing profits and in identifying potential customers who are most likely to respond to a new product. More details of these techniques and their applications in other fields will be discussed in the next blog.
Dr. J Alamelu Mangai
Computer Science Engineering,
Presidency University, Bengaluru – 59