Content-Based Image Retrieval (CBIR) in Remote Clinical Diagnosis and Healthcare

1. Introduction

Medical Image Processing (MIP) is a major ally in e-health and telemedicine, enabling rapid diagnosis with visual, quantitative and analytical assessment (Fessler, 2009). Remote care can reveal subtle changes that indicate the progression of a therapy. Health facilities now have images from plenty of sources which leads to multidimensional images (2D, 3D, 4D, etc.), and multimodality images. For instance, Alzheimer’s disease evaluation still uses behavioral and cognitive tests along with MRI and PET scans of the entire brain (Datta et al., 2008; Deselaers, 2014; Lew et al., 2006). Diverse image collections offer the chance to improve evidence-based diagnosis, administration, teaching, and research. There is a necessity for proper methods to search those collections for images that have similarity in some sense. Statistical bias can be reduced as discoveries are assessed without direct patient contact like quicker and more objective assessment of the effects of anticancer drugs.

Content-Based Image Retrieval (CBIR) is an image search framework that complements the usual text-based retrieval of images through visual features, such as color, shape, and texture as search criteria.

CBIR can be applied to multidimensional image retrieval, multimodality health data, and the recuperation of unusual datasets.

CBIR Systems (CBIRSs) can be divided into two classes: Narrow Domain Applications (NDA) and Broad Domain Applications (BDA).

Medical Imagery Retrieval, Finger Print Retrieval, and Satellite Imagery Retrieval are types of NDA. These applications have small variance of content; target specific sources of knowledge; have homogeneous semantics; are likely to have some sort of ground truth; their content description is more objective; may have some control of scenes and sensors; involve limited interactivity; use quantitative evaluation; have tailored/data-driven architectures; are medium-sized; use object recognition techniques most of the times; and consider specific invariances with model-drivel tools.

Photo collections and Internet content are examples of BDA. They have high content variance; tackle generic sources of knowledge; have heterogeneous semantics; do not have any sort of ground truth; their content description is more subjective; no control of scenes and sensors; involve pervasiveness and interactivity; use qualitative evaluation; have modular/iteration-driven architectures; range from large to extremely large sizes; use information retrieval techniques most of the times; consider tools that are perceptual/cultural, in addition to general invariants.

Besides facilitating visual/automatic diagnosis and decision making, images can help real-time remote consultation and screening, store-and-forward examinations, home care assistance and overall patient surveillance. These applications involve a great deal of data that require web-based and other telemedicine structures. When it comes to the use of information systems in healthcare, some ideas must be clarified. Next, some relevant definitions are given (Oh, et al., 2005; Embi & Payne, 2009):