Abstract
34
Objectives: bjectives Patient misidentification accident in medical examination, for example a wrong patient is scanned and the wrong images are registered on PACS, has been one of the big problems in hospitals. On the other hand, recently, image analysis using convolutional neural network (CNN), which is one of machine learning methods, is rapidly becoming popular as a part of artificial intelligence (AI) system for medical imaging. CNN is one of the deep learning techniques and is known to be feasible to image analysis by recognizing complex visual patters in a similar way to human perceptions. The misidentification accidents could be prevented if AI predicts patient characteristics (sex, body weight, etc) automatically from image itself and alerts when detecting mismatches. Therefore, in this research using CNN, we aimed to develop a system that predicts patient sex from FDG PET-CT images. Methods This retrospective study included sequential 10,000 patients who underwent whole-body FDG PET-CT with either a Siemens Biograph 64 scanner (N=8,342) or a Philips GEMINI TF-64 scanner (N=1,658) at our institute between January 2015 and August 2017. Scans of same patients were excluded. Maximum intensity projection (MIP) images (matrix size, 168 × 168) were generated by rotation by 10 degrees (19 images per person). CT images were not used in this study. CNN was conducted for classifying patients by sex. CNN was trained and validated using the randomly selected 70% images while the data of the remaining 30% images were used for test
Purpose: The process was repeated 5 times to calculate the accuracy. This experiment was performed under the following environment: OS, Windows 10 pro 64 bit; CPU, intel Core i7-6700K; GPU, 2 × NVIDIA GeForce GTX 1070 8GB; Framework, Keras 2.0.2 and TensorFlow 1.3.0; Language, Python 3.5.2; CNN, original CNN (Convolution layer, 5; Maxpooling layer, 4); Optimizer, Adam. Results The 10,000 patients included male (N=5014) and female (N=4986) patients almost equally. Thus, a total of 190,000 images were provided, and each category consisted of approximately 95,000 images. The CNN process spent ~80 minutes for training each fold dataset and <1 minute for prediction. When images for test purpose were given to the learned model, the accuracy was up to 94.5±3.0%. Furthermore, for patient-based classification, the patient sex was predicted from 19 MIP images based on majority rule. Roughly all patients were correctly categorized, resulting in overall accuracy of 97.6%.
Conclusions: The current data suggested that the CNN-based sex prediction system successfully classified the patients, which may be useful to prevent patient misidentification accidents in daily clinical settings.