Sumeet Dua and Mohit Jain

Fig. 3.6. Example of decision tree after assigning event probability.

Since the weight of “Everyone” is higher, everyone is included for the diagnostic examination. The above example illustrates how decision trees can help an ophthalmologist make complex clinical decisions and diagnose patients.

3.3.Use of Information Technologies for Diagnosis in Ophthalmology

Information technology helps healthcare professionals to build decision support systems, such as diagnosis, therapy planning, and monitoring.4 The components of decision support systems are image acquisition, data mining, and graphical user interface (GUI). There are four ways in which images can be acquired, as explained below.

1.Slit-lamp cameras are used to capture the details of the eye structure by changing the parameters or width of the beam.

2.P2, a digital, integrated platform for ocular imaging, combines the functions of the slit lamp and BIO with a CCD camera and acquires 3D images from the front and back of the eye without dilation. The 3D images accurately depict the current eye condition and will help improve diagnostic accuracy.

3.A RetCam can be used by nurses in neonatal intensive care units (NICUs) to take images.

4.Optical coherence tomography (OCT) creates high quality, micrometerresolution, 3D images from within optical scattering media.5

98

Computational Decision Support Systems and Diagnostic Tools

Data mining, or knowledge discovery, can be used to gather and utilize information in healthcare. Because it handles multiple data entries, it is ideal for ophthalmology, where clinics may have a variety of patients with similar diseases but wide-ranging symptoms and personal circumstances.

3.3.1. Data Mining in Ophthalmology

Data mining consists of five steps: preprocessing, feature extraction, feature selection, clustering, and classification.6 Image preprocessing consists of three steps. If images have different brightness or intensity, the clinician must normalize all the images to the same brightness level, or it will be difficult to differentiate between images of healthy and/or diseased eyes. This process can be completed by using an image histogram. An image histogram is a pictorial representation of the distribution of intensities of the image through which one can increase or decrease the intensity or brightness of the image as per the requirement for better viewing. An image histogram helps to create uniform brightness for each image and improve the quality of the image.

As with brightness, every image should be the same size. It is difficult to perform matrix operations, such as addition or subtraction, on images that are not the same size. Images that are not the same size can be normalized by padding, i.e. adding 0s that do not affect picture information to lengthen the shorter images. Padding does not distort the image by adding external information. Noise is undesirable information that has to be removed from the image. Gaussian noise,7 quantization noise,8 and film grain9 are methods that can be used to remove noise.

Feature extraction is a form of dimensionality reduction. When the input data is large and redundant, then the feature extraction algorithm will divide large datasets into smaller sections or into feature vectors. For feature extraction, a system first reads an image and extracts the features or key points of the image that will be represented as unique features. The feature vectors help to distinguish between healthy and diseased eyes. Feature extraction can be performed by extracting the features of the image using global or regional feature extraction. For global feature extraction, the system will take the entire image as input and extract the feature vectors. The global feature extraction can be performed using algorithms, such as scale invariant

99

Sumeet Dua and Mohit Jain

feature transform10 and principal component analysis.11 For regional feature extraction, the system will divide the image into several equal-size parts. For each part, the system will extract the features and then merge the extracted features into one feature vector. From the vectors, the system can extract the regional features by using algorithms, such as gray-scale co-occurrence matrix12 and grid-based segmentation.13 The feature vector will become the output of these algorithms for the next phase.

Feature selection is also known as attribute selection or as feature reduction. Feature selection aids in selecting important or relevant features. By selecting important features, the clinician can speed up the learning process or remove the effect of the curse of dimensionality. Feature selection can be performed by feature ranking, where the clinician will get the rank of the features or the attributes, and, based on the rank, the clinician can select the features. Methods by which the clinician can rank our features include the ChiSquared attribute,14 the GainRatio attribute,15 and the infogain attribute.16 The clinician can use tools such as Weka, Orange, and Rapidminer, as well for feature ranking.17

Clustering is the process of grouping together objects that are similar to one another but are dissimilar to objects in other clusters or groups. Once the clinician has selected the features, he or she can give a class label to the features or to the category in which they belong, i.e. healthy eyes or the diseased eyes. However, if the clinician does not know the class labels or the category title, then we use clustering to learn the class labels or the category to which group the feature belongs. The quality of a cluster will depend on the following three factors: high intra-class similarity and low inter-class similarity, the similarity measure used by the method and its implementation, and the cluster’s ability to discover hidden patterns.

Once the clinician has the class labels, then he or she can begin classification. A classification algorithm is used to predict categorical or discrete values based on the training data or the data the clinician gets after feature selection or clustering. There are two steps for this process: model construction and model usage. In model construction, the clinician builds or trains a model according to the training data, and in model usage, the clinician classifies future or unknown objects or estimates the accuracy of the classification. In ophthalmology, the clinician determines how many images have been correctly classified to the healthy eyes and diseased eyes classes.

100