Abstract

Abstract. In this article we study the use of SPECT perfusion imaging for the diagnosis of Alzheimer's disease. We present a classifier based approach that does not need any explicit knowledge about the pathology. We directly use the voxel intensities as features. This approach is compared with three classical approaches: regions of interests, statistical parametric mapping and visual analysis which is the most commonly used method. We tested our method both on simulated and on real data. The realistic simulations give us total control about the ground truth. On real data, our method was more sensitive than the human experts, while having an acceptable specificity. We conclude that an automatic method can be a useful help for clinicians.

The purpose of this study was to compare sensitivity for detection of pulmonary nodules in MDCT scans and reading time of radiologists when using CAD as the second reader (SR) respectively concurrent reader (CR). Four radiologists analyzed 50 chest MDCT scans chosen from clinical routine two times and marked all detected pulmonary nodules: first with CAD as CR (display of CAD results immediately in the reading session) and later (median 14 weeks) with CAD as SR (display of CAD markers after completion of first reading without CAD). A Siemens LungCAD prototype was used. Sensitivities for detection of nodules and reading times were recorded. Sensitivity of reading with CAD as SR was significantly higher than reading without CAD (p < 0.001) and CAD as CR (p < 0.001). For nodule size of 1.75 mm or above no significant sensitivity difference between CAD as CR and reading without CAD was observed; e.g., for nodules above 4 mm sensitivity was 68% without CAD, 68% with CAD as CR (p = 0.45) and 75% with CAD as SR (p < 0.001). Reading time was significantly shorter for CR (274 s) compared to reading without CAD (294 s; p = 0.04) and SR (337 s; p < 0.001). In our study CAD could either speed up reading of chest CT cases for pulmonary nodules without relevant loss of sensitivity when used as CR, or it increased sensitivity at the cost of longer reading times when used as SR.

No abstract

This work aims at providing a tool to assist the interpretation of SPECT images for the diagnosis of Alzheimer's Disease (AD). Our approach is to test classifiers, which uses the intensity values of the images, without any prior information. Such a classifier is built upon a training set, containing images with two different labels (AD patients and normal subjects). It will then provide a classification for any new unknown image. The main problem to be handled is the small number of available images compared to the large number of features (here the image's voxels): the so-called small sample size problem. We evaluate here the ability of two linear classifiers to correctly label a set of 79 images. Our experiments show promising results. They also show that image classification based on intensity values only is possible and might be used for other applications as well.