Saad et al. present an excellent overview of the problems associated with the interpretation of medical images.
Traditionally, medical images may be acquired by computerized tomography (CT) and magnetic resonance imaging (MRI) scanners. The images may be merged together for a patient. Ultimately, the anatomical structures that are shown in these images have to be isolated and identified by someone. This is a very time-consuming process.
The authors suggest that rather than focusing on a single, isolated patient, it can be helpful to present the user of the medical imaging data with supplementary information that places the patient in the context of the population. In other words, does a patient conform to or deviate from a large population of healthy people?
The authors use a framework that consists of atlas construction and segmentation uncertainty interaction. The segmentation interaction is guided by knowledge learned from expert-segmented images of a training population. This is an interesting approach because it assists a user of medical imaging data. The goal is to be able to identify suspicious regions (for example, tumors) and to correct the misclassified results.
In order to reach the framework of atlas construction and segmentation uncertainty interaction, the authors give an excellent discussion of the problems inherent in identifying suspicious regions on a medical image and their approach for resolving these problems. They present a series of examples that are clear and can be quickly grasped.
Since this paper is aimed at a nonmedical audience, it glosses over some of the problems associated with accurately acquiring medical images. Omitting these problems simplifies the presentation and expedites the authors’ approach for describing a solution to their audience.
This is an excellent article that could be of interest to people working within an imaging environment. Although it focuses on medical imaging, the approaches taken by the authors could be easily applied to a general, nonmedical audience.
