In recent years, the rapid evolution of artificial intelligence (AI) has brought about a revolution in medical research. In particular, the application of AI technology to medical imaging is expanding rapidly, with image-to-image translation technique gaining significant attention. Image-to-image translation technique allows for a wide range of applications, such as converting between different imaging modalities and removing artifacts. It is expected to open up new perspectives that go beyond the traditional framework of medical imaging. Using image-to-image translation models, it’s possible to generate synthetic PET from MRI images, or convert images with artifacts to those without, potentially contributing to improved diagnostic accuracy and optimization of treatment plans. In this article, we introduce two papers we published applying image-to-image translation technique in the field of neuroradiology: a study on generating synthetic methionine PET using MRI, and a study on producing Digital Subtraction Angiography (DSA) without misregistration artifacts.

Deep learning has dramatically improved the prediction accuracy of various medical data analyses, including image classification. However, for predictions related to critical decisions such as medical diagnoses, it is essential not only to consider average accuracy but also to quantify the uncertainty of individual predictions and evaluate their reliability. This review provides an overview of probabilistic predictions through machine learning, their calibration, proper scoring rules, and methods to distinguish different types of uncertainties. Furthermore, for the situation with label uncertainties as commonly observed in blood cell image classifications, methods to evaluate and enhance uncertainty predictions based on label frequency are introduced.

We attempt to classify cancer tumors into several feature groups by counting holes based on topology. In our opinion, image features of cancer tumors depend on several factors such as tumor cells, necrosis, and angiogenesis. When analyzing tumors topologically, we assume that the image features of cancer can be decomposed into three structural components (three-dimensional space): connected components, holes, and cavities. In this paper, we would like to describe a topology-based “hole analysis” including why cancer features are converted into numbers with topology, the theory of the hole analysis, why cancer features are related to topology, and several applications of the hole analysis.

Advancements in post-genome researches have deepened our understanding of molecular biology and enable the detailed subcategorization of diseases. In the cancer treatment, patients are stratified based on genetic information of their cancer, and precision medicine tailored to individual cancer characteristic is been performed. Radiogenomics is the study of estimating cancer’s somatic variants from images, and radioproteomics focuses on estimating protein activity from images. If these studies are successful, it may eliminate invasive biopsies, offer a more cost-effective way, and propose the most suitable treatment options. This paper describes the current state of radiogenomics and radioproteomics researches.

This review introduces some decision support techniques to assist diagnosis and treatment using convolutional neural networks (CNN), one of the deep learning technologies. As technologies to assist physicians in diagnosis and treatment, our research group has developed survival prediction using CT images of lung cancer patients and postoperative urine abstinence prediction using robotic surgery videos of prostate cancer patients. In addition, to assist radiological technologist in their imaging work, our research group has developed a CNN-based assessment method for artificial knee joint radiographs, and a method for converting non-contrast enhanced CT images to contrast enhanced CT images using CycleGAN. Above techniques were first realized using CNNs, and they will play an important role in decision making.

Risk assessment plays a crucial role in reducing mortality and optimizing healthcare resource utilization in cardiac care units (CCU). However, the criteria for admitting cardiac patients to CCU often rely on subjective judgments by cardiologists. Therefore, our study aimed to evaluate risk by predicting patient prognosis at discharge using multimodal learning, incorporating electrocardiography (ECG), a routine examination conducted upon CCU admission, and clinical data. We included 892 survivors and 289 non-survivors, who were admitted to the CCU between 2008 and 2020, with prognosis at discharge as the endpoint. The 12-lead ECGs obtained at admission were converted into images, and ECG features were extracted using a convolutional neural network. These features were then combined with clinical data, comprising 22 items, to predict the prognosis at discharge using gradient boosting. Additionally, the contribution of each feature was calculated using SHapley Additive exPlanations. The 10-fold cross-validation demonstrated an improved accuracy, with an AUC of 0.889, compared to using ECG and clinical data independently. Notably, the machine learning model exhibited a particular focus on myocardial troponin and eGFR as influential factors. These findings suggest that our proposed method holds promise for enhancing risk assessment in CCU patients.

Prostate cancer progresses more slowly than other cancers and can be cured with appropriate treatment. Recently, the use of minimally invasive robot-assisted radical prostatectomy (RARP) in surgery, one of the treatment methods, has made it possible to return to society earlier than with conventional methods. However, the impact of urinary incontinence, which is a side effect, on patient’s daily lives cannot be ignored, and prediction of postoperative urinary continence is expected to assist in the selection of treatment options. In this study, magnetic resonance images taken at the time of diagnosis at multiple facilities were preprocessed, and then classified into good and bad cases of urinary continence using 3D convolutional neural network (3DCNN) and machine learning methods. The degree of urinary continence was defined as good or bad based on the number of urinary pads changed per day at 3 months postoperatively. As a result, accuracy for good urinary continence was 83.7%, accuracy for bad urinary continence was 67.1%, and Balanced accuracy was 75.4%. These results indicate that the classification accuracy is higher than that of previous studies on 2D images, suggesting that this method may be an effective technique for predicting postoperative urinary continence in prostate cancer patients.

Radiomics has been established to support treatmentdecision making in precision medicine. The radiomic features (RFs) used for prediction must be stable with respect to thevariety of imaging conditions. The purpose was to evaluate the stability of RFsextracted from T1-weighted MR images (T1WIs) using different imaging conditions. Two types of stabilityevaluation (SE) phantoms that canbe used for contrast (C-SE) andresolution (R-SE) assessment werecreated. The T1WIs of each phantom were acquired. Regarding the imagingparameters, the number of excitations (NEX), matrix size (MS),and repetition time (TR) werevaried. A total of 837 RFs were extracted from each T1WI acquired withdifferent parameters. Stability was evaluated using the coefficient ofvariation (CV). The criterion ofstability was employed as the CV < 0.05. The percentage of stable featuresin the C- and R-SE phantoms against individually changes in imaging conditionswere 30.8% and 31.1% for the TR, 34.8% and 39.1% for the NEX, and 35.6% and 38.5% for the MS respectively. Moreover, the percentage of stable features forchanging all conditions were 21.0% for the C-SE phantom and 22.1% for the R-SEphantom. Stable features were found in MR images against changes in imagingconditions.