Brain perfusion SPECT with an automated quantitative tool can identify prodromal Alzheimer's disease among patients with mild cognitive impairment.

OBJECTIVE: To improve diagnosis of early Alzheimer's disease (AD), i.e., prodromal AD, by an automated quantitative tool combining brain perfusion single-photon emission computed tomography (SPECT) images and memory tests scores in order to be applied in clinical practice. PATIENTS AND METHODS: In this prospective, longitudinal, multi-centric study, a baseline (99m)Tc-ECD perfusion SPECT was performed in 83 patients with memory complaint and mild cognitive impairment (MCI). After a 3-year follow-up, 11 patients progressed to Alzheimer's disease (MCI-AD group), and 72 patients remained stable (MCI-S group), including 1 patient who developed mild vascular cognitive impairment. After comparison between the MCI-S and MCI-AD groups with a voxel-based approach, region masks were extracted from the statistically significant clusters and used alone or in combination with Free and Cued Selective Reminding Test (FCSRT) scores for the subject's categorization using linear discriminant analysis. Results were validated using the leave-one-out cross-validation method. RESULTS: Right parietal and hippocampal perfusion was significantly (p<0.05, corrected) decreased in the MCI-AD group as compared to the MCI-S group. The patients' classification in the MCI group using the mean activity in right and left parietal cortex and hippocampus yielded a sensitivity, specificity, and accuracy of 82%, 90%, and 89%, respectively. Combination of SPECT results and FCSRT free recall scores increased specificity to 93%. CONCLUSION: The combination of an automated quantitative tool for brain perfusion SPECT images and memory test scores was able to distinguish, in a group of amnestic MCI, patients at an early stage of AD from patients with stable MCI.

Detection of melanoma from dermoscopic images of naevi acquired under uncontrolled conditions.

BACKGROUND AND OBJECTIVE: Several systems for the diagnosis of melanoma from images of naevi obtained under controlled conditions have demonstrated comparable efficiency with dermatologists. However, their robustness to analyze daily routine images was sometimes questionable. The purpose of this work is to investigate to what extent the automatic melanoma diagnosis may be achieved from the analysis of uncontrolled images of pigmented skin lesions. MATERIALS AND METHODS: Images were acquired during regular practice by two dermatologists using Reflex 24 x 36 cameras combined with Heine Delta 10 dermascopes. The images were then digitalized using a scanner. In addition, five senior dermatologists were asked to give the diagnosis and therapeutic decision (exeresis) for 227 images of naevi, together with an opinion about the existence of malignancy-predictive features. Meanwhile, a learning by sample classifier for the diagnosis of melanoma was constructed, which combines image-processing with machine-learning techniques. After an automatic segmentation, geometric and colorimetric parameters were extracted from images and selected according to their efficiency in predicting malignancy features. A diagnosis was subsequently provided based on selected parameters. An extensive comparison of dermatologists' and computer results was subsequently performed. RESULTS AND CONCLUSION: The KL-PLS-based classifier shows comparable performances with respect to dermatologists (sensitivity: 95% and specificity: 60%). The algorithm provides an original insight into the clinical knowledge of pigmented skin lesions.

Differential automatic diagnosis between Alzheimer's disease and frontotemporal dementia based on perfusion SPECT images.

OBJECTIVE: Alzheimer's disease (AD) and frontotemporal dementia (FTD) are among the most frequent neurodegenerative cognitive disorders, but their differential diagnosis is difficult. The aim of this study was to evaluate an automatic method returning the probability that a patient suffers from AD or FTD from the analysis of brain perfusion single photon emission computed tomography images. METHODS AND MATERIALS: A set of 116 descriptors corresponding to the average activity in regions of interest was calculated from the images of 82 AD and 91 FTD patients. A set of linear (logistic regression and linear discriminant analysis) and non-linear (support vector machines, k-nearest neighbours, multilayer perceptron and kernel logistic PLS) classification methods was subsequently used to ascertain diagnoses. Validation was carried out by means of the leave-one-out protocol. Diagnoses by the classifier and by four physicians (visual assessment) were compared. Since images were acquired in different hospitals, the impact of the medical centre on the diagnosis of both the classifier and the physicians was investigated. RESULTS: Best results were obtained with support vector machine and partial least squares regression coupled with k-nearest neighbours methods (PLS+K-NN), with an overall accuracy of 88%. PLS+K-NN was however considered as the best method since performances obtained with leave-one-out cross-validation were closer to whole-database learning. The performances of the classifier were higher than those of experts (accuracy ranged from 65 to 72%). Physicians found it more difficult to diagnose the images from centres other than their own, and it affected their performances. CONCLUSIONS: The performances obtained by the classifier for the differential diagnosis of AD and FTD were found convincing. It could help physicians in daily practice, particularly when visual assessment is inconclusive, or when dealing with multicentre data.