Supervised learning such as deep learning has applied to various tasks and achieved high accuracy in biomedical image analysis. However, it is required to prepare sufficient labeled data to learn discriminative features for robust recognition. It often requires significant effort by biomedical experts to annotate images for various objects, imaging modality and types of disease. In this paper, I introduce current works about unsupervised, semi-supervised, and weakly-supervised learning, which enables to reduce the annotation costs. I then introduce our researches on these learning problems.
Unsupervised image segmentation is an important technique in various research fields such as medical image processing. Basically, unsupervised image segmentation is based on some hand-crafted features and clustering pixels in a way that takes into account feature similarity and spatial continuity. In contrast, the authors proposed a method that applies unsupervised learning of convolutional neural networks (CNNs) to image segmentation. The proposed CNN estimates to which cluster each pixel in the input image belongs, as in a general supervised image segmentation task. However, it does not require any supervisory signals of pixel labels or network pre-training, and the network is trained only after the target image is input. In this paper, we describe the conventional basics of such unsupervised image segmentation, as well as the authorʼs proposed method using deep learning.
In this paper, we introduce methods for estimating the 6DoF pose of objects using point clouds. This technique has two problem settings: instance-level pose estimation, which estimates the pose of a specific 3D model, and category-level pose estimation, which estimates the pose of an unseen object. We introduce some representative examples of both problem settings. In addition, we introduce our recent work on category-level pose estimation by self-supervised learning.
Unsupervised anomaly detection is an important task in medical image analysis. On the other hand, there is a deep relationship between generative deep model and unsupervised anomaly detection. In this paper, firstly the generative models are described, and it is followed by an introduction of several generative deep models. Then, some topics are given in which unsupervised anomaly detection was applied to medical image analysis tasks. Finally, a newest method of self-supervised representation learning-based anomaly detection is described.
Optical spectroscopy provides functional and morphological information of living tissues based on optical absorption and scattering properties. Spectral diffuse reflectance imaging can be achieved with a simple and robust optical system, and has been widely applied to in vivo measurements and imaging of biological tissues. In the present article, various spectral imaging techniques based on diffuse reflectance and their applications to biomedical imaging are described.
Brain age estimation based on machine learning from MRI images has been attracting attention as a biomarker of the progression of brain degeneration in such as Alzheimerʼs disease and temporal lobe epilepsy. However, it is known that measurement bias has been existed in the volume due to differences in image quality of MRI obtained from different magnetic field strength, manufacturer and model, which has been reported to significantly reduce the generalizability of machine learning. In this study, we used ComBat harmonization method, which uses a general linear model and empirical Bayesian estimation to correct the measurement bias between scanner and site, which improved the generalizability and accuracy of brain age estimation across different.
C-shaped PET, in which a part of the ring is open, may expand the application of PET, such as an add-on PET insert for an existing MRI system. However, the measurable projection data of C-shaped PET are truncated, and strong artifacts are generated in its reconstructed images. Therefore, we have proposed C-shaped Compton-PET, C-shaped PET with a scatterer insert on the opposite side of the ring open part to compensate for the truncation by applying the Compton camera technique. This study conducted a 3D imaging simulation using the Geant4 Monte Carlo toolkit to demonstrate C-shaped Compton-PET. We modeled C-shaped Compton-PET geometry. A scatterer (15 cm inner diameter) was arranged in an arc and inserted into a C-shaped PET (20 cm inner diameter and an opening angle of 115 degrees). We simulated a uniform cylindrical phantom with a radius of 10 cm and a height of 15 cm, and reconstructed images were evaluated quantitatively. As a result, the percentage ratio of total pixel values in the cylinder region against the entire FOV increased from 89% to 95% by the scatterer insert. The cylinder shape was reproduced more accurately than the C-shaped PET. In conclusion, we demonstrated that C-shaped Compton-PET is effective for reducing image artifacts.
At present, the research and development of computer-aided diagnosis (CAD) is being conducted in various medi cal fields. Because of the difficulty in separating the large intestine from peripheral organs, little development has been carried out in the area of CAD that employs CT. Recently, due to improved endoscopic precision and the widespread use of CT colonography, research has been underway into such areas as applying CAD to discover colorectal cancer. However, endoscopy and CT colonography require advance preparation. CT colonography in particular is difficult to conduct in some cases because of stress placed on the colon when the colon is inflated by injecting carbon dioxide gas. In this study, research and development was conducted into a method for segmenting, the large-intestine region from plain abdominal CT images captured during an abdominal examination. The proposed multi-step extraction method extracts the optimal shape based on the features by binarizing the data while gradually varying the threshold value from high to low. Using this method makes it possible to acquire images for use in CAD development without necessitating colon-stressing CT colonography. The concordance rate of 82% between the segmentation results of this method and Ground-truth demonstrate this methodʼs segmentation performance.
Deep learning has realized image classification methods by a data-driven mechanism. It can be said that the relationship between the input data and the correct answer was learned from a large amount of data. When tackling an image classification problem using this principle, it is necessary to collect a large amount of data on rare events in order to deal with rare events, but this can be an impossible task in reality. Anomaly detection is a unique approach that realizes the separation of rare events based on an idea of defining the range of normal features using a large amount of normal data and determining the degree of abnormality by deviation from it. The image feature extraction method using an autoencoder and its evaluation method was described in this course.