Ultrasound images contain more noise than those obtained from imaging modality such as CT (computed tomography) and MRI (magnetic resonance imaging). The development of AI for image diagnosis support using artificial intelligence (AI) is still in its early stages. The authors have focused on fetal cardiac ultrasound screening, which is one of the most difficult ultrasound examinations. The authors have proposed various methods based on the consideration that improving explainability is essential for the widespread use of ultrasound AI in medical practice. The authors propose explainable representations such as barcode-like timeline and graph chart. Furthermore, the authors explain how doctors can improve their screening abilities by using both representations. Finally, the authors introduce their prospects of ultrasound AI research.

Brain image data analysis, based on voxel values, plays a crucial role in diagnosing and treating brain diseases. Recently, artificial intelligence techniques have been applied to this field, but challenges remain in improving processing speed and ensuring the interpretability of results. This paper focuses on scoring methods using dimensionality reduction. Appropriately derived scores can serve as brain image biomarkers, potentially lowering computational costs while enhancing explanatory power. Anatomical standardization, used as a preprocessing step, enables the application of matrix decomposition techniques. The multi-supervised sparse component analysis proposed here extends conventional matrix factorization by efficiently reducing large-scale brain image data through stepwise linear transformations. The inverse transformation helps identify relevant anatomical regions, thereby improving interpretability. Additionally, we demonstrate how to eliminate unnecessary variation by selectively inverting components of the multi-supervised method, using specific analytical examples.

A criterion for grading follicular lymphoma is defined by the World Health Organization (WHO) based on the number of centroblasts and centrocytes within the field of view. However, the WHO criterion is not often used in clinical practice because it is impractical for pathologists to visually identify the cell type of each cell and count the number of centroblasts and centrocytes. In this paper, based on the widespread use of digital pathology, we make it practical to identify and count the cell type by using image processing and then construct a criterion for grading based on the number of cells. Here, the problem is that labeling the cell type is not easy even for experienced pathologists. To alleviate this problem, we build a new dataset for cell type classification, which contains the pathologists’ confusion records during labeling, and we construct the cell type classifier using complementary-label learning from this dataset. Then we propose a criterion based on the composition ratio of cell types that is consistent with the pathologists’ grading. Our experiments demonstrate that the classifier can accurately identify cell types and the proposed criterion is more consistent with the pathologists’ grading than the current WHO criterion.

Digital images in the field of pathology are known as whole slide images (WSIs), which are high-resolution scans of glass specimens captured by WSI scanners. Due to their large size, WSIs cannot be directly input into machine learning models; instead, small image patches extracted from WSIs are typically used as input. However, annotating tumors within these images requires significant effort from pathologists, making it difficult to assign accurate labels to each patch. As a result, it is common in digital pathology to treat the problem as one in which a single class label is assigned to an entire WSI. In such weakly supervised settings, multiple instance learning (MIL) has proven effective. In particular, attention-based MIL (ABMIL) has become a widely adopted model, as it allows for visualization of attention weights to highlight regions within a WSI that contribute to the model’s decision. This makes ABMIL a promising approach for explainable AI in digital pathology. This paper provides an overview of ABMIL and its applications in the context of explainable AI.

Nearly 30 years ago, the author developed a multi-scale local structure analysis method for extracting line-like (blood vessels, airways, etc.), sheet-like (cortex, articular cartilage, etc.), and blob-like (tumors, lymph nodes, etc.) structures from 3D medical images. This method is still utilized in the field of biomedical image processing today. This paper looks back on the background and circumstances under which this analysis method was born during the dawn of collaborative research between medicine and information science, reflecting on the research environment at the time, interactions with the advisor, and the process of trial and error. Initially, there were cautious views regarding the direction of the research, but through sustained efforts, gradual progress was achieved, eventually leading to a turning point. Additionally, while this research was conducted in a U.S. laboratory, it was significantly influenced by medical image research in Japan at the time.