1.1 Background of the CAD System
Computer-aided diagnosis (CAD) has become one of the major research approaches in medical imaging science. In this book, the motivation and philosophy for the early development of CAD systems are presented together with the glaucoma CAD system. With CAD systems, radiologists and doctors use the computer output as a “second opinion,” and then make the final decisions. CAD is a concept established by taking into account the importance of physicians and computer applications equally, whereas automated computer diagnosis depends on computer algorithms only (Doi et al., 2017).
“Today, the role of computer-aided diagnosis is expanding beyond screening programs and toward applications in diagnosis, risk assessment and response to therapy.” (Giger, 2010)
The computer-aided detection (CAD) or computer-aided diagnosis (CAD) is the computer software or application that helps doctors to take decisions swiftly (Doi, 2017), (Li and Nishikawa, 2015). Medical imaging contains data in images evaluated by doctors to analyze abnormality and diagnose diseases. Analysis of imaging in the medical field is a very important approach because imaging is the basic modality to diagnose diseases at the earliest opportunity, and image acquisition is almost always an external operation that does not cause harm. Imaging techniques like MRI, X-ray, endoscopy, ultrasound, etc., if acquired with high energy will provide a good-quality image, but they will harm the human body; however, images that are taken with less energy will sometimes be poor in quality and have a low contrast. CAD systems are used to improve the image quality, helping to detect the medical image’s abnormality correctly and diagnose different diseases (Chen et al., 2013). Computer-aided diagnosis (CAD) technology includes multiple science aspects, such as artificial intelligence (AI), computer vision, and medical image processing. The main application of CAD systems is in finding abnormalities in the human body and diagnosing diseases as a second opinion. For example, the detection of tumors is a typical application because if they are diagnosed early in basic screening, it will help prevent the spread of cancer (Giger, 2000).
1.2 Objectives and Significance of the CAD System
The main goal of CAD systems is to identify signs of abnormality at the earliest possible opportunity where a human professional might fail to notice them. For example, in mammography, the identification of small lumps in dense tissue, finding architectural distortion, and prediction of mass type as benign or malignant by its shape, size, etc.
CAD is usually restricted to mark the visible abnormalities in images, whereas CAD helps to evaluate the abnormalities identified in the images at an earliest disease stage. For example, it highlights abnormalities in RNFL, breaking the ISNT rule, showing texture and color abnormalities in the retina as the earliest signs of glaucoma. This helps the radiologist to draw their conclusion. Though CAD has been used for over 40 years, it still does not reach its expected outcomes. We agree that CAD cannot act as a substitute for a real doctor, but it definitely helps radiologists become better decision makers. It plays as a second opinion and final interpretative role in medical diagnosis.
1.4 Previous Studies of CAD Systems
Halalli et al. (2017) proposed a CAD system to detect breast cancer where he addressed detection steps such as pre-processing, segmentation, feature extraction, and classification. He proposed a CAD used for identification of subtle signs for use in breast cancer detection and classification.
Hiroshi et al. (2008) proposed a CAD system for the early detection of (1) cerebrovascular diseases using brain MRI and MRA images by detecting lacunar infarcts, unruptured aneurysms, and arterial occlusions; (2) other types of ocular diseases such as glaucoma, diabetic retinopathy, and hypertensive retinopathy using retinal fundus images; and (3) breast cancers, using ultrasound 3-D volumetric whole breast data using the breast masses detection.
Na Young et al. (2014) proposed a study to evaluate the diagnostic performance of the CAD system in full-field digital mammography for detecting breast cancer when used by a dedicated breast radiologist (BR) and a radiology resident (RR), in order to detect who could most benefit from a CAD application. They found that CAD was helpful for the BRs to improve their diagnostic performance and for RRs to improve their sensitivity in a screening setting, and concluded that CAD could be essential for radiologists by decreasing reading time and improving diagnostic performance.
The CAD accuracy was generally better in the BR group than in the RR group, but sensitivity was better with CAD use in both groups, from 81.10% to 84.29% in the BR group, and from 75.38% to 77.95% in the RR group. The most improvement in disease diagnosis was observed in the BR group, whereas in the RR group sensitivity improved but specificity, PPV, and NPV did not. The main advantage of using the CAD system was shortened time in both the BR and RR groups, from 111.6 minutes to 94.3 minutes for BR, and 135.5 minutes to 109.8 minutes for RR, which was more significant for the RR group than it was for the BR group.
Godoy et al. (2013) proposed a study to evaluate the impact of CAD on the identification of sub solid and solid lung nodules on thin- and thick-section CT.
They used 46 chest CT examinations with ground-glass opacity (GGO) nodules; CAD marks computed using thin data were evaluated in two phases, and for 155 nodules they found (mean, 5.5 mm; range, 4.0–27.5 mm) – 74 solid nodules, 22 part-solid (part-solid nodules), and 59 GGO nodules – CAD stand-alone sensitivity was 80%, 95%, and 71%, respectively, with 3 false-positives on average (0–12) per CT study. Reader(thin) + CAD(thin) sensitivities were higher than reader(thin) for solid nodules (82% vs. 57%, p < 0.001), part-solid nodules (97% vs. 81%, p = 0.0027), and GGO nodules (82% vs. 69%, p < 0.001) for all readers (p < 0.001). Respective sensitivities for reader(thick), reader(thick) + CAD(thick), reader(thick) + CAD(thin) were 40%, 58% (p < 0.001), and 77% (p < 0.001) for solid nodules; 72%, 73% (p = 0.322), and 94% (p < 0.001) for part-solid nodules; and 53%, 58% (p = 0.008), and 79% (p < 0.001) for GGO nodules. For reader(thin), false-positives increased from 0.64 per case to 0.90 with CAD(thin) (p < 0.001) but not for reader(thick); false-positive rates were 1.17, 1.19, and 1.26 per case for reader(thick), reader(thick) + CAD(thick), and reader(thick) + CAD(thin), respectively.
Automating mass chest screening for tuberculosis (TB) requires segmentation and texture analysis in chest radiograph images (Bram, 2001). Several rule-based schemes, pixel classifiers, and active shape model techniques for segmenting lung fields in chest radiographs are described and compared. An improved version of the active shape model segmentation technique, originally developed by Cootes and Taylor from Manchester University, UK, is described that uses optimal local features to steer the segmentation process and outperforms the original method in segmentation tasks for several types of medical images, including chest radiographs and slices from MRI brain data. In order to segment the posterior ribs in PA chest radiographs, a statistical model of the complete rib cage is constructed using principal components analysis and a method is described to fit this model to input images automatically. For texture analysis, an extension is proposed to the framework of locally orderless images, a multi-scale description of local histograms recently proposed by Koenderink and Van Doorn from Utrecht University, the Netherlands. The segmentation and texture analysis techniques are combined into a single method that automatically detects textural abnormalities in chest radiographs and estimates the probability that an image contains abnormalities. The method was evaluated on two databases. On a 200-case database of clinical chest films with interstitial disease from the University of Chicago, excellent results were obtained in the area under the ROC Curve 50.99. The results for a 600-case database from a TB screening program are an encouraging area under the ROC Curve 50.82.
The aim of the Marco et al. (2004) study was to evaluate diagnostic accuracy provided by different statistical classifiers on a large set of pigmented skin lesions grabbed by four digital analyzers located in two different dermatological units. Experimental Design: images of 391 melanomas and 449 melanocytic nevi were included in the study. The methodology was based on a linear classifier to identify a threshold value for a sensitivity of 95%, therefore, a K-nearest-neighbor classifier, a nonparametric method of pattern recognition, was constructed using all available image features and trained for a sensitivity of 98% on a large dataset of lesions. The obtained result on independent test sets of lesions for the linear classifier and the K-nearest-neighbor classifier produced a mean sensitivity of 95% and 98%, and a mean specificity of 78% and of 79%, respectively. They then suggested that computer-aided differentiation of melanoma from benign pigmented lesions obtained with DB-Mips is feasible and, above all, reliable. In fact, the same instrumentations used in different units provided similar diagnostic accuracy. Whether this would improve early diagnosis of melanoma and/or reduce unnecessary surgery needs to be demonstrated by a randomized clinical trial.