Medical advances depend on technological advances. In practice, in many clinical situations, it is often possible to improve work without using such high-tech. Asking what you feel daily is a step toward improvement. To achieve this purpose, it is necessary to feel the inconvenience using your five senses on site. Instead of making a custom-made product based on the idea of one person, we must create a product that many people want and use. You have to figure out how much a product is needed. Think carefully about the features that your product requires. The product should be as simple as possible. There is a manufacturing concept called “BioDesign” advocated by Stanford University. We should be aware of the principles of failure : Fail often, Fail early, and Fail cheap. In this paper, I will present my experience of making medical products through cardiovascular surgery.

Non-small cell lung cancer (NSCLC) tumors are categorized into three histological subtypes (adenocarcinoma :AC, squamous cell carcinoma : SCC and large cell carcinoma : LC). Histological classification of NSCLC affects the decision making of treatment policies. However, histological subtypes identified from specimens sampled by a single biopsy occasionally differ from those from surgical resection. For increasing the classification accuracy, we aim to develop an automated approach for classifying NSCLC cases into major two histological subtypes, AC and SCC by using two machine learning techniques ; support vector machine (SVM) and random forest (RF) with radiomic features. After calculating intraclass correlation coefficients (ICCs) for investigating reproducible radiomic features, we extracted 31 stable features using CT images of 155 NSCLC patients, and applied five feature selection methods. A leave-one-out cross validation was performed and model parameters were optimized to maximize the area under receiver operating characteristic curves (AUCs). Finally, the optimized models were applied to 50 NSCLC patients. The most robust combination of the feature selection and machine learning techniques was considered as the SVM classification model using the signature constructed by RF, which achieved the highest AUC of 0.7577. The proposed method based on radiomic features could classify NSCLC into AC and SCC.

For hepatocellular carcinoma, complicated stratifications and treatment methods have being studied. However,the 5-year survival rate of hepatocellular carcinoma in stage I is 60.4%, which is lower than that for breast cancer or lung cancer. The purpose of this study is to investigate whether radiomic feature and gene expression are useful for stratifying the prognosis of hepatocellular carcinoma. From a public database TCGA-LIHC（The Cancer Genome Atlas Liver Hepatocellular Carcinoma）, contrast-enhanced CT images and gene expression levels of 19 cases, including 11 cases were death 2 years later, were selected for this study. 5 radiomic features and 5 genes were selected by Lasso, and hepatocellular carcinomas were stratified using principal coordinate analysis and clustering. In addition, we analyzed the relationship between radiomic feature and gene expression using canonical correlation analysis. Experimental results showed that gene expression was more useful for stratifying hepatocellular carcinomas than radiomic feature. However, since the canonical correlation analysis can be applied to search for radiomic features that are complementary to gene data, we believe that the proposed method is an important technique in considering the division of roles between image and gene examinations in the future.

In this paper, we propose an unsupervised dynamic positron emission tomography (PET) image denoising scheme using a deep image prior with anatomical information as convolutional neural network input. The proposed conditional DIP method is an unsupervised deep learning technique with no need to prepare any prior training of image datasets including high- and low-quality image pairs, and only learns using a single data pair of a target dynamic PET image and the anatomical information of a patient’s own magnetic resonance image. A numerical simulation is performed using a three-dimensional brain phantom with ¹⁸F-FDG kinetics. Results show that using the anatomical information, the proposed conditional DIP method yielded improved image quality and quantitative performance.