Abstract
Objectives: Our aim was to design a semantic association feature to represent brain PET images to improve the accuracy of neuroimaging classification.
Methods: The low-level features including the mean index, Fisher index and the difference of Gaussian-based features (lesion volume index, lesion metabolism index and lesion contrast index), were firstly extracted from PET data. We then built a dictionary using k-means clustering. Then, a Probabilistic Latent Semantic Analysis (PLSA) was used to learn the latent topics based on the word co-occurrence among all images so that each image was represented as a mixture over these topics, each of which was a probability distribution upon the extracted words. A Canonical Correlation Analysis (CCA) was then used to capture the latent associations by making the topic distributions of each image pair as correlated as possible, and represent the images as mixtures of the transformed topics in this image pair context.
Results: Our method was tested data in the Alzheimer’s disease Neuroimaging Initiative (ADNI) database. There were 331 subjects annotated as cognitively normal (77), having mild cognitive impairment (169) and having Alzheimer’s disease (85). The classification rates using the semantic association features were compared to the rates using only the low-level features based on two different classifiers, i.e., k-nearest neighbour (k-NN) and support vector machine (SVM). Our method had higher classification rates (60% with SVM, 65% with kNN) than the low-level features (54% with SVM, 46% with kNN), representing improved neuroimaging classification.
Conclusions: Our method extracts semantic associations to bridge the gap between the low-level features and high-level descriptions and enhances neuroimaging classification.
Methods: The low-level features including the mean index, Fisher index and the difference of Gaussian-based features (lesion volume index, lesion metabolism index and lesion contrast index), were firstly extracted from PET data. We then built a dictionary using k-means clustering. Then, a Probabilistic Latent Semantic Analysis (PLSA) was used to learn the latent topics based on the word co-occurrence among all images so that each image was represented as a mixture over these topics, each of which was a probability distribution upon the extracted words. A Canonical Correlation Analysis (CCA) was then used to capture the latent associations by making the topic distributions of each image pair as correlated as possible, and represent the images as mixtures of the transformed topics in this image pair context.
Results: Our method was tested data in the Alzheimer’s disease Neuroimaging Initiative (ADNI) database. There were 331 subjects annotated as cognitively normal (77), having mild cognitive impairment (169) and having Alzheimer’s disease (85). The classification rates using the semantic association features were compared to the rates using only the low-level features based on two different classifiers, i.e., k-nearest neighbour (k-NN) and support vector machine (SVM). Our method had higher classification rates (60% with SVM, 65% with kNN) than the low-level features (54% with SVM, 46% with kNN), representing improved neuroimaging classification.
Conclusions: Our method extracts semantic associations to bridge the gap between the low-level features and high-level descriptions and enhances neuroimaging classification.
Original language | English |
---|---|
Article number | 2029 |
Number of pages | 1 |
Journal | The Journal of Nuclear Medicine |
Volume | 55 |
Issue number | Supplement 1 |
Publication status | Published - May 2014 |
Externally published | Yes |