Since the beginning of modern medicine, advances in diagnostic and therapeutic equipment, advances in diagnostic agents, and advances in therapeutic agents have made great progress in medicine. Artificial intelligence (AI),which is currently being developed and implemented at a great pace, will play a major role in the future development of medical care and medicine. Keio University Hospital has established a medical AI center in the school of Medicine and has been adopted by the Cabinet Office for AI hospital projects, which is increasing the momentum for introducing new medical care utilizing AI and IT technologies in the hospital. This paper introduces the efforts of AI hospital project at Keio University Hospital.

There have been many studies on the use of AI for the diagnosis of COVID-19 pneumonia. Although these AIs have high performance, there are several points that need to be considered when using them in clinical practice. In this paper, we comprehensively reviewed the AI studies for COVID-19 pneumonia that were registered in PubMed until December 2020. The performance and problems of these AIs were presented.

Radiation pneumonitis (RP) is a radiotherapy dose-limiting toxicity for non-small-cell lung cancer. Prior studies have proposed clinical and dosimetric factors related to RP. This review paper describes about a radiomics-based predictive model for RP after lung cancer stereotactic body radiation therapy (SBRT) on pretreatment planning computed tomography (CT) images. For the RP prediction models, 275 non-small-cell lung cancer patients consisted of 245 training (22 with grade ≥2 RP) and 30 test cases (8 with grade ≥2 RP) were selected. A total of 486 radiomic features were calculated to quantify texture patterns within lung volumes. Ten subsets consisting of all 22 RP cases and 22 or 23 randomly selected non-RP cases were created from the imbalanced dataset of 245 training patients. For each subset, signatures were constructed, and predictive models were built using the least absolute shrinkage and selection operator (LASSO) logistic regression. An ensemble averaging model was built by averaging the RP probabilities of the 10 models. The best model areas under the receiver operating characteristic curves (AUCs) calculated on the training and test cohorts were 0.871 and 0.756, respectively. The radiomic features calculated on pretreatment planning CT images could be predictive imaging biomarkers for RP after lung cancer SBRT.

Novel coronavirus pneumonia has found in late 2019, and it has spread around the world. COVID-19 pandemic has changed the way of our lives tremendously. WHO reported 170 million people infected, and the Ministry of Health,Labour and Welfare of Japan has reported 0.75 million people infections in Japan on June 7 ^th . Medical image diagnosis using X-ray photograph or X-ray CT are practical tools for COVID-19 infections since they enable us to take images of entire of the lung. Especially, chest CT images have significant advantages to observe the inside the lung on axial slices and are easy to diagnose ground-glass opacity regions. This paper shows image analysis techniques for COVID-19 cases by introducing our laboratory’s work on COVID-19 CT image analysis. Furthermore, we discuss image analysis, network image database and computation facilities for the COVID-19 pandemic.

Our research group has been developing computer-aided diagnosis systems for diffuse lung disease (DLD) on CT images. Because appearance of DLD is rich in diversity, the diagnosis may be different depending on doctors. On the other hand, because the diagnosis affects a method of treatment, accurate diagnosis is required. In this paper, we introduce a method for segmenting lung regions with DLD and a method for segmenting DLD regions themselves for each type of DLD. The results of these methods, the Dice coefficient in the former method and the classification rate for the latter method were 0.98 and 85%.

This paper reviews the application of deep learning model to the automatic diagnosis of COVID-19 on chest x- ray images. Among various deep learning models proposed for the automatic diagnosis of COVID-19, COVID-NET, CV19- Net, and the authors' model are introduced. The source code is publicly available for these three models, and the datasets of chest x-ray images are also available for two of them. It is expected that these publicly available source codes and datasets of diagnostic models will be useful for the research on diagnostic models of COVID-19 and other diseases. Finally, future work on the author's diagnostic model for COVID-19 is presented.

Many AI-CADs have been developed to assist physicians in the diagnosis of lung cancer using CT images. However, insufficient number of data and bias have hindered improving the performance of AI-CADs. In this review article,we describe the generation of lung nodule images by generative adversarial networks and its application to AI-CADs as one of the solutions to this challenge.

CAD and radiomics have been used to develop AI systems in the field of radiology. CAD systems support the detection and differential diagnoses of various diseases. These systems estimate the present disease state from current images. Radiomics supports medical care after a disease is detected by predicting prognostic and therapeutic effects. Therefore, radiomics differs from CAD in that it predicts a future state from a current image. This paper describes the current state of radiomic research for lung cancer.

Convolutional neural networks (CNNs) intrinsically requires large-scale data whereas chest X-ray (CXR) images tend to be data/annotation-scarce, leading to over-fitting. Therefore, based on our development experience and related work,this paper thoroughly introduces tricks to improve generalization in the CXR diagnosis : how to (i) leverage additional data,(ii) augment/distillate data, (iii) regularize training, and (iv) conduct efficient segmentation. As a development example based on such optimization techniques, we also feature LPIXEL's CNN-based CXR solution, EIRL Chest Nodule, which improved radiologists/non-radiologists' nodule detection sensitivity by 0.100/0.131, respectively, while maintaining specificity.

Despite ongoing efforts on improving machine learning model performance, capturing model uncertainty remains a major challenge in medical imaging, an important subject that has been largely overlooked by the community. This work borrows standard model calibration approaches and empirically demonstrates their effectiveness on medical imaging triage with labels automatically extracted by different natural language processing techniques. We demonstrate both the strength and weakness of three different calibration methods using two sets of NLP labels. The tests are conducted on human-labelled ground truth. Although all methods yield comparable results, our proposed approach further improves AUCs when paired with a strong NLP model that generates smooth labels.

Children have a high risk of infection of pneumonia, and Chest X-ray images are used in pneumonia diagnosis. However, the diagnosis of pneumonia varies among physicians. We developed an automated detection scheme of pneumonia in pediatric chest X-ray images. This method has two steps that is extraction of lung regions to obtain the analysis regions and detection pneumonia with radiomic features. Mask R-CNN including object detection and semantic segmentation was used to extract the lung regions. Regions of interest (ROI) of 32 × 32 pixels were set along the shape of the extracted lung regions, and radiomic features were calculated in each ROI. The ROIs were classified into normal and pneumonia based on the radiomic features using deep neural network. As for the training of network, chest X-ray images of ChestX-ray 8 database were used and the images acquired at Tokyo Metropolitan Children's Medical Center were used for testing. As the result of extraction of lung regions, average of dice index was 96.1%. The accuracy of pneumonia classification based on radiomic features was 80.5%. Therefore, high accuracy was obtained even if the training and test images were taken from different country, age and hospitals.

With the development of object detection research, computer-aided detection research has made great progress in pulmonary nodule detection in these years. Especially for chest radiographs (CXRs), convolutional neural networks (CNNs) have achieved encouraging performance on several public datasets. However, most of them are supervised learning methods and rely on annotation of nodule location. Because of the lack of annotated information, these methods may not be applicable for many large-scale CXR datasets. In this paper, we suggest a weakly-supervised method for pulmonary nodule detection for addressing this issue. To alleviate the influence from the absence of annotated data during training, we train a CNN for nodule classification and calculate the gradient-weighted Class activation mapping (Grad-cam) of CNN to visualize the attention of the model. Without nodule location, it is hard to generate the candidate of nodule from Grad-cam at high accuracy. Thus, a self-calibrated block is then developed to improve the detection by calculating the Grad-cam with different transformations. Extensive experiments on Chest X-ray 14 dataset and JSRT dataset show considerable performance improvement compared with other weakly-supervised methods for pulmonary nodule detection.

Respiratory disease is a serious illness that accounts for three of the top ten causes of death in the world, and approximately eight million people died worldwide each year. Early detection and early treatment are important for the prevention of illness due to these diseases. Currently, auscultation is performed for the diagnosis of respiratory diseases,however there is a problem that quantitative diagnosis is difficult. Therefore, in this paper, we propose a new automatic classification method of respiratory sounds to support the diagnosis of respiratory diseases on auscultation. In the proposed method, respiratory sound data is converted into a spectrogram image by applying the short-time Fourier transform. Then,we apply HPSS (Harmonic/Percussive Sound Separation) algorithm to the respiratory sound spectrogram to separate it into a harmonic spectrogram and a percussive spectrogram. The three generated spectrograms are used for classification of respiratory sounds by CNN (Convolutional Neural Network) and SVM (Support Vector Machine) classifiers. Our proposed method obtained superior classification performance compared to the case without applying HPSS and satisfactory results are obtained.

Though it is common practice to examine patients for abdominal CT with the tube voltage of 120 kV, there is an expectation that the same image quality could be achievable by adopting lower tube voltages such as 100 kV with reduced patient dose, based on the fact that the iodine contrast is higher with lower X-ray energies. If this is really the case,the dose reduction will depend on patient size too. Thus the purpose of this work was to investigate patient size dependency of dose reduction efficiency with low-kV CT. Under the approval of institutional review board, we compared abdominal scan images taken with 120 kV and 100 kV, 150 patients each, in terms of dose-weighted contrast to noise ratio (CNRD),since it is the value of not dependent on the mAs value applied. As for dose parameter, we adopted size specific dose estimate (SSDE) instead of volume CT dose index (CTDIvol) which is being used widely, since CTDIvol does not represent patient size dependency of dose correctly. As a result in terms of CNRD of liver and spleen contrast, 100 kV outperformed 120 kV with statistical significance throughout all the patient sizes investigated. The difference was size dependent, being larger with smaller patient size.

In tomosynthesis imaging, acquired from the general X-ray apparatus, the center of rotation (COR) of the X-ray tube must be set before image acquisition. However, there is a lack of literature on the impact of the height of the COR setting on the quality of the projection image of tomosynthesis. This study aims to examine the effect of the COR settings on the modulation transfer function (MTF) of the projection image. The edge method was employed to calculate the MTF. The tungsten edges were positioned at intervals of 20 mm, from 0 mm to 160 mm above the table. One-shot radiography and projection images were obtained from each tungsten edge position. Two types of COR settings were defined at the height of the edge (Variable COR) and 80 mm from the table (Fixed COR). The MTF of the focal spot (MTF_focal) was calculated from the one-shot image, and the quotient of the MTF, from the projection image and the MTF_focal. This quotient will be the MTF after rejecting the effect of the focal spot (MTF_tomo). In the Variable COR setting, the MTF_tomo were decreased by rising the height of the edge device. We also confirmed that the MTF_tomo. degraded when the distance from the COR to the tungsten edge device exceeded ±40 mm in the Fixed COR setting.