Nowadays, positron emission tomography (PET) scan is focused in the field of disease diagnosis. In order to obtain a clear image in the PET scan, it is necessary to increase the S/N ratio, which leads to an increase in exposure dose at the time of observation. For this reason, it is desired to increase the S/N ratio of the image while suppressing the exposure dose. In this research, we applied a method combining two noise reduction methods to this problem. First, we used a noise reduction method using dictionary learning for the sinogram representation. The second used a noise reduction method for the real image representation based on the regularization approach. As a result, it was found that an approach combining two noise reduction methods is more effective than the conventional method.

In recent years, as chemotherapy has advanced, it is important to accurately diagnosis the histological type (adenocarcinoma, squamous cell carcinoma and small cell carcinoma). Pathologists diagnose not only images but also the patientʼs clinical background. In this study, we aimed to develop automated classification scheme of lung cancer type by combining liquid-based cytological (LBC) images and electronic medical record. First, image features were extracted from LBC images using deep convolutional neural network (DCNN). Subsequently, patient clinical data (smoking status etc.) were collected, and dimension compression was performed by principal component analysis (PCA). Image features and patient clinical data corresponding to cytological images were given to the classifier. Finally, classification result of 3 histological categories was obtained. In the experiments, the proposed method was applied to 149 cases (Adenocarcinoma; 50, Squamous cell carcinoma; 51, Small cell carcinoma; 48) and evaluated via 3-fold cross-validation. As a result of experiments, the classification accuracy of the cytological image alone using DCNN was 82.9%. When the image feature and patient basic data (age, gender, smoking status etc.), or image feature and tumor markers were given to the support vector machine (SVM), the classification accuracy was improved. These results indicate that the proposed method may be useful for histological classification of lung cancer.

To reduce image artifacts originating from statistical noise and photon attenuation in emission CT, iterative image reconstruction algorithms have been widely used. However, these methods require high computational costs. Furthermore, in recent years, attenuation correction methods using anatomical CT/MRI images have been investigated, but these methods tend to require rather complicated computations. In this paper, we propose use of convolutional neural network (CNN) for the image correction in the emission CT. In this method, we input an FBP reconstructed PET/SPECT image with no smoothing and no attenuation correction into the CNN, which is processed by the CNN to obtain a final corrected image. We also investigated to input a pair of an FBP degraded image and a corresponding CT image (multimodality image) into the CNN. The learning of CNN was performed by using a set of images constructed through simulating image formation process of the emission CT. The simulation results demonstrate that the CNN-based method enables noise reduction and attenuation correction simultaneously in a short time. Also, the results demonstrate that inputting a CT image in addition to an emission CT image further improves the correction of low frequency artifact.

Computer-aided diagnosis (CADx) systems based on deep learning have been actively researched in recent years, and it has been reported to exhibit their high performance. The CADx systems for benign and malignant discrimination exhibit a similar tendency. They generally convert input medical images into likelihood of their benignancy and malignancy. On the other hand, when medical doctors explain the diagnosis to patients, they need to explain not only the likelihood of malignancy of the lung nodule but also basis of the diagnosis. In this paper, we proposed a CADx system which provides medical doctor with likelihood of the existence of the medical image finding related to lung cancer CNN (convolutional neural network) for obtaining the likelihood of the medical image findings, and NN (neural network) for obtaining the likelihood of the malignancy of the lung nodule. As a result of evaluating the performance of the proposed system using 55 benign and 120 malignant nodules, the discrimination rate was 79.02±8.43 [%].