Various approaches have been taken to detect anomalies, with certain particularities in the medical image scenario, linked to other terms: content-based image retrieval, pattern recognition, classification, segmentation, outlier detection, image mining, as well as computer-assisted diagnosis, and computeraided surgery. This chapter presents, a review of anomaly detection (AD) techniques and assessment methodologies, which have been applied to medical images, emphasizing their peculiarities, limitations and future perspectives. Moreover, a contribution to the field of AD in brain computed tomography images is also given, illustrated and assessed.
Most AD algorithms for medical image analysis are profoundly influenced by the specific image datasets used and by the medical or biological task. Figure 1 shows this diversity at a glance. Most reported studies have dealt with detection of tumors in digital mammography (Huang, 2004; Selvi, 2005; Wei, 2005; Peng, 2006; Chiracharit, 2007; Ikedo, 2007; Karnan, 2007), lung CT images (Minhas, 2005; Sluimer, 2006), and brain magnetic resonance (MR) images (Gering, 2003; Prastawa, 2004; Lee, 2005; Benamrane, 2006; Menze, 2006; Shinkareva, 2006; Bouix, 2007; Ekin, 2007), but many others can be mentioned.
Key Terms in this Chapter
Measures of Performance: Measures used to evaluate the performance of AD algorithms, by using images with available ground truths. Most of them are based on the coincidences (true positives and true negatives) and not coincidences (false positives and false negatives) between regions detected/classified by algorithms under assessment and the corresponding regions in the ground truth.
Window/Level Adjustment: Mapping of portions of the image dynamic range to the dynamic range of the display monitor. For instance, a 12 bit CT image should be re-scaled for brain matter analysis with a window/level adjustment around 35 ± 35 Hounsfield Units (i.e. gray-levels between 1000 and 1070). This is the base of the variable-bin-size histogram approach in this work.
Anomaly: Deviation or departure from the normal or common order, form or rule; one that is peculiar, irregular, abnormal or difficult to classify. In image analysis, anomalies are unknown targets, which are relatively small and with low probability of occurrence. Tumors, micro-calcifications, and vascular irregularities are examples of anomalies in the medical image analysis framework.
Content-Based Image Retrieval: Process of retrieving images from databases based on its real visual contents (features of texture, shape, and color) by using signal processing, pattern recognition and computer vision methods.
Ground Truth: The image gold standard for assessing detection/classification algorithms. This term was originally used to designate the true information gathered in ground to evaluate remote sensing techniques like aerial photographs or satellite imagery, but it has been generalized to other scenarios.
Anomaly Detection (AD) Systems: Systems used to detect anomalies. They can be developed with no prior knowledge of the data, or modeling both normality and anomalies, or modeling only normality. In the medical imaging context, the third approach is the best suited. AD systems can be based on statistical methods, neural networks, or machine learning.
Segmentation: The partitioning of digital images into different regions, which group elements (pixels or voxels) with similar feature values.
Imaging Modalities: Different physical principles involved in the acquisition of an image. In the medical imaging context, we can mention: photography, endoscope, microscopy, electrical impedance tomography, ultrasound-based systems, X rays, CT, MRI and fMRI, MRSI, SPECT, PET and PET/CT.