Cervical lesion classification has gained the attention of researchers because of its importance in the mitigation and treatment of cervical cancer. Most machine learning approaches have addressed the task under the single and fully labeled assumption; however, such assumptions do not reflect the nature of cervical lesions in a clinical setting. In this study, we adapt a semi-supervised learning algorithm on a partially labeled multi-label cervigram dataset by treating it as positive-unlabeled. We simultaneously trained a classifier, and a propensity model to simulate the clinical bias in labeling lesions. Results were compared to that of a supervised learning model, showing improvements on several performance metrics. Gradient-weighted class activation mapping also showed better learning focus.

In Japan, mammography (MMG) alone or in combination with palpation is used for primary breast cancer screening. In MMG imaging, it is difficult to distinguish breast cancer in “dense breast” with many mammary glands and little fat, or in subjects in their 30s or younger, who have well-developed mammary gland, because the white areas in the images are more likely to appear white. This is particularly true in the case of breast masses, which have lower sensitivity and specificity. If the accuracy of classifying benign and malignant shadows of breast cancer masses in MMG-imaging images could be improved, it would lead to earlier detection of breast cancer. In this study, we developed an image quality improvement process that simultaneously improves the granularity and contrast of images to enhance the performance of CAD for benign and malignant classification using MMG images. As a result, the accuracy was improved to 89% when the image quality improvement process was applied, compared to 85% when the MMG images without image quality improvement process were used for benign-malignant classification. We also analyzed the factors that the learned AI focused on in the image when making discrimination in the CAD process for benign-malignant classification and report the results. In general, the screening by using screening MMG alone or in combination with palpation, and differential diagnosis is made by using diagnostic MMG, ultrasonography, and MRI at the second stage of detailed examination. This study proposes that improved image quality will enable differentiation even with screening MMG and increase the possibility of early detection of breast cancer.

Over the last decade, in pathology numerous reports agree that performance of artificial intelligence (AI) with convolutional neural networks in diagnostic imaging has exceeded that of humans. The consequent increase of expectations on AI has raised a problem of accountability: while human doctors are always responsible for their diagnosis, the same does not apply to an AI doctor. In particular, the decision process of the AI is concealed within a black box that does not provide the reason for which a certain diagnosis is reached. To address this problem, known as explainability, we propose the following method. The input image for the diagnosis contains many local features. Among them, features which are often seen in the lesion area and rarely in the normal area have a high content of information and should be highlighted as information-rich and likely evidence of the lesion. When we define “Information Density” as the total amount of information that features in each sector of the image carry, the sectors with large “Information Density” are identified as those belonging to the lesion. The results of this method on CAMELYON 16 images are consistent with the pathologistʼs diagnosis.

Rheumatoid arthritis causes systemic joint destruction and significantly interferes with quality of life for patients. The diagnosis of rheumatoid arthritis is made by blood tests and X-ray imaging. Radiologist can easily monitor the disease progression using X-ray images, however it has the problems of increased burden on physicians and lack of reproducibility. Therefore, the development of computer aided diagnosis systems that use the results of computer analysis as a second opinion is expected. In this paper, we propose an image registration method for phalanges CR image to develop computer aided diagnosis system that detects the progression of rheumatoid arthritis from past and current images of the same patient using a temporal subtraction method. In this study, we develop a new method based on geometric matching CNN, which can be optimized for each image pair, to precisely align the phalanges region. We applied our proposed CNN method to 560 pairs of synthetic phalanges images and obtained TP of 99.26％ and FP of 0.79％ respectively, confirming the effectiveness of the proposed method in comparison with previous methods.

In classification problems, Cross-Entropy Loss can be used to separate features in the feature space. On the other hand, Contrastive Learning is possible to obtain useful representations by learning features so that the features of the same class are close and those of the different classes are far from each other. In this paper, we focus on Supervised Contrastive Learning (SCL), which uses label information to embed features more appropriately within the framework of supervised learning, and applying it to the task of classifying the opacities of chest CT images. We found that the classification accuracy was improved by 8-18% in the four validation patterns performed in terms of adaptation to each of the two different domains (Hospital 1 cases and Hospital 2 cases) and adaptation across domains. Furthermore, by visualizing the obtained features using t-SNE, we confirmed that the groups of classes are created by SCL clearly compared with the method with Cross-Entropy Loss.

It is necessary to extract surgical instruments from laparoscopic images in order to improve the safety of laparoscopic surgery using a surgery support system. It is reported that the segmentation accuracy can be improved by using color and depth information. In this paper, we propose a U-Net based image segmentation network using the estimated depth by pix2pix as well as color for improving the accuracy. We conducted experiments using 4-fold cross validation with 1,800 images in the MICCAI challenge dataset, and confirmed that the proposed method achieved the average IoU of 88％ and the average Dice coefficient of 93%. The proposed method reduced the excessive extraction and improved the extraction accuracy by using the estimated depth information as well as color information.

Pathologists visually observe hematoxylin and eosin (HE) stained images under a microscope to perform pathological diagnosis. However, it is not possible to sufficiently diagnose future recurrence after surgery by judging shape using HE stained specimens alone, it is difficult to properly formulate a treatment policy for patients. In order to accurately identify tumor recurrence, this study proposes a method of automatically identifying future recurrence in a pathological image by calculating features of RGB and LLL (Image obtained by copying the luminance image L to 3 channels) components and by using the parallel structure of feature extractors. The method consists of three steps: 1. Features of recurrence presence or absence of surgically resected lung adenocarcinoma IB are extracted from RGB and LLL components of HE stained image using a convolutional neural network (CNN), 2. a classifier is created so that those features identify recurrence presence or absence of lung adenocarcinoma IB by using the CNN, 3. the recurrence presence or absence of lung adenocarcinoma is judged by using the classifier. The experimental results using digital images of pathological tissue specimens of lung adenocarcinoma IB show improved identification accuracy.

The mechanism of edge artifacts in point spread function (PSF) reconstruction was investigated by a simulation study. A two-dimensional (2D) numerical phantom comprises 184 mm diameter disk background and 4.6-30 mm diameter disk signal with contrast 9 (expressed as signal/background −1). Parallel beam projection data were blurred by detector response of 1D Gaussian function with 5 mm full width at half maximum (FWHM). Projection matrix only considers detector response function, ensuring that the edge artifacts are not masked or affected by other components. Image reconstruction was performed using ordered subset expectation maximization (OSEM) without regularization and with regularization by generalized Gaussian function. Without regularization, the object-specific modulation transfer function (OMTF), defined as the ratio of power spectrum of original and reconstructed images, was ~1 up to 1.6 cycles/cm, lower than the kernel frequency support of the detector response function (1/FWHM: 2.0 cycles/cm) and then decayed rapidly. This frequency response is similar to the original image and was smoothed using a Butterworth filter of order 4 and a cutoff frequency of 2.94 cycles/cm, producing the edge artifacts in image space. In PSF reconstruction, the edge artifacts are generated because the frequency response has the shape of low-pass Butterworth filter, lacking high-frequency components.