Detection of Alzheimer’s disease using ECD SPECT images by transfer learning from FDG PET

• Published in: Annals of nuclear medicine Vol. 35; no. 8; pp. 889 - 899
• Main Authors: Ni, Yu-Ching, Tseng, Fan-Pin, Pai, Ming-Chyi, Hsiao, Ing-Tsung, Lin, Kun-Ju, Lin, Zhi-Kun, Lin, Wen-Bin, Chiu, Pai-Yi, Hung, Guang-Uei, Chang, Chiung-Chih, Chang, Ya-Ting, Chuang, Keh‑Shih
• Format: Journal Article
• Language: English
• Published: Singapore Springer Singapore 01.08.2021; Springer Nature B.V
• Subjects: Accuracy; Aged; Aged, 80 and over; Alzheimer Disease - diagnostic imaging; Alzheimer's disease; Brain; Brain - diagnostic imaging; Cognition; Cognitive ability; Cysteine - analogs & derivatives; Datasets; Deep Learning; Feature extraction; Female; Fluorine isotopes; Fluorodeoxyglucose F18; Humans; Image Processing, Computer-Assisted - methods; Imaging; Male; Mathematical models; Medical imaging; Medicine; Medicine & Public Health; Model accuracy; Neurodegenerative diseases; Neuroimaging; Nuclear Medicine; Organotechnetium Compounds; Original Article; Parameters; Perfusion; Positron emission; Positron emission tomography; Positron-Emission Tomography - methods; Radioisotopes; Radiology; Sensitivity; Single photon emission computed tomography; Tomography; Tomography, Emission-Computed, Single-Photon; Training; Transfer learning; Alzheimer’s disease; ECD SPECT images; Transfer learning
• ISSN: 0914-7187; 1864-6433; 1864-6433
• DOI: 10.1007/s12149-021-01626-3

Objective To develop a practical method to rapidly utilize a deep learning model to automatically extract image features based on a small number of SPECT brain perfusion images in general clinics to objectively evaluate Alzheimer's disease (AD). Methods For the properties of low cost and convenient access in general clinics, Tc-99-ECD SPECT imaging data in brain perfusion detection was used in this study for AD detection. Two-stage transfer learning based on the Inception v3 network model was performed using the ImageNet dataset and ADNI database. To improve training accuracy, the three-dimensional image was reorganized into three sets of two-dimensional images for data augmentation and ensemble learning. The effect of pre-training parameters for Tc-99m-ECD SPECT image to distinguish AD from normal cognition (NC) was investigated, as well as the effect of the sample size of F-18-FDG PET images used in pre-training. The same model was also fine-tuned for the prediction of the MMSE score from the Tc-99m-ECD SPECT image. Results The AUC values of w/wo pre-training parameters for Tc-99m-ECD SPECT image to distinguish AD from NC were 0.86 and 0.90. The sensitivity, specificity, precision, accuracy, and F1 score were 100%, 75%, 76%, 86%, and 86%, respectively for the training model with 1000 cases of F-18-FDG PET image for pre-training. The AUC values for various sample sizes of the training dataset (100, 200, 400, 800, 1000 cases) for pre-training were 0.86, 0.91, 0.95, 0.97, and 0.97. Regardless of the pre-training condition ECD dataset used, the AUC value was greater than 0.85. Finally, predicting cognitive scores and MMSE scores correlated ( R 2  = 0.7072). Conclusions With the ADNI pre-trained model, the sensitivity and accuracy of the proposed deep learning model using SPECT ECD perfusion images to differentiate AD from NC were increased by approximately 30% and 10%, respectively. Our study indicated that the model trained on PET FDG metabolic imaging for the same disease could be transferred to a small sample of SPECT cerebral perfusion images. This model will contribute to the practicality of SPECT cerebral perfusion images using deep learning technology to objectively recognize AD.