PT  - JOURNAL ARTICLE
AU  - Takuro Shiiba
AU  - Akihiro Takaki
TI  - Comparison of diagnostic performance of deep convolutional neural network using fine-tuning and feature extraction on dopamine transporter single photon emission tomography images
DP  - 2019 May 01
TA  - Journal of Nuclear Medicine
PG  - 1219--1219
VI  - 60
IP  - supplement 1
4099  - http://jnm.snmjournals.org/content/60/supplement_1/1219.short
4100  - http://jnm.snmjournals.org/content/60/supplement_1/1219.full
SO  - J Nucl Med2019 May 01; 60
AB  - 1219Objectives: Recently, deep convolutional neural network (DCNN) is widely applying to research on medical imaging. It is well knownthat the transfer learning efficient using fine-tuning (FT) or feature extraction (FE) for DCNN is higher than that of developing network from the scratch. However, the diagnostic performance of DCNN using FT and FE in nuclear medicine images has not been investigated. Thus, we compared diagnostic performance between FT and FE on 123-I FP-CIT dopamine transporter single photon emission computed tomography (SPECT). Methods: A data set of 213 normal control (NC) and 432 Parkinson’s disease (PD) 123-I FP-CIT SPECT images. These images were divided into 451 and 194 for training and test, respectively. The DCNN of both FT and FE used AlexNet. In the FT, the last three layers of AlexNet were replaced with a fully connected layer, a softmax layer, and a classification output layer to classify into two groups. Training images for FT were augmented with randomly preprocessing such as rotation, translation, and reflection. In the FE training, features were extracted with fully connected layer ‘FC7’ of AlexNet, and fit a support vector machine. After completion of training, the diagnostic accuracy of FT and FE by using test images were compared, and the receiver operating characteristic (ROC) analysis was performed. Results: The test diagnostic accuracy of FT and FE were following: sensitivity, 96.9% and 90.0%; specificity, 85.9% and 75.0%; positive predictive value, 93.3% and 88.0%; and negative predictive value, 93.2% and 78.7%. The area under the ROC curves of FT and FE in test were 0.974 and 0.897, respectively. Conclusions: This study demonstrated that the classification performance of FT and FE for diagnosis of PD and NC. FT indicated more accurate diagnostic performance than that of FE. Therefore, FT has a high potential to construct automatic diagnosis system for 123-I FP-CIT dopamine transporter SPECT images.