Tatsuhito Hasegawa

Diffusion-based dataset distillation has recently emerged as a promising paradigm for condensing large-scale datasets into compact synthetic sets. By leveraging pretrained generative priors, these methods can produce realistic class-conditional samples more efficiently than traditional matching-based approaches. However, most existing diffusionbased methods still adopt a rigid “Generate-and-Use” strategy, where the generated samples are directly treated as the final distilled set under a fixed images-per-class budget. Such a design tightly couples candidate generation with final budget allocation, which may result in redundant waste of the limited budget or insufficiently informative samples. In this paper, we propose “Pool-Select-Refine”, a two-stage framework for allocation-aware generative dataset distillation. First, instead of directly using a fixed number of generated samples, we construct an over-complete candidate pool and select a compact subset under the target budget. Second, we refine the selected samples in latent space using soft-label supervision derived from the teacher model, improving semantic alignment while preserving the generative prior. This design explicitly decouples generation, selection, and refinement, enabling more effective use of the distillation budget. Experiments on large-scale and fine-grained image classification benchmarks show that the proposed framework delivers consistent gains over diffusion-based baselines. The results suggest that introducing a curation stage before refinement is a simple yet effective way to improve diffusion-based dataset distillation.

Temperature is a critical environmental factor influencing the physiology and behavior of ectothermic animals, yet conventional methods for evaluating thermal tolerance in fish rely on subjective manual observation of loss of equilibrium (LOE), limiting experimental throughput and introducing observer bias. Here, we developed an automated temperature tolerance evaluation system integrating DeepLabCut-based pose estimation with custom image processing algorithms to objectively quantify the timing of LOE during thermal stress tests. Our system incorporated region partitioning and color transformation preprocessing to improve keypoint detection accuracy, followed by a classification model combining ResNet34-based frame features with keypoint coordinates to objectively determine the timing of LOE without manual observation. Validation against manual annotation showed that the automated system achieved an accuracy comparable to the natural variability between trained investigators, and outperformed naive human observers, supporting its validity as an objective and reproducible alternative to manual scoring. Using this system, we characterized cold and heat tolerance across six medaka strains ( Oryzias latipes : d-rR/TOKYO, HB11A, OK-Cab, HO5 and HdrR-II1; O. sakaizumii : HNI-II). Cold and heat tolerance assessment revealed inter-strain variation, with HdrR-II1 among the most cold- and heat-tolerant strains and HNI-II the least tolerant of both cold and heat stress. We further evaluated cold tolerance in medaka-related species ( O. sinensis , O. cabaranensis , O. curvinotus , O. luzonensis , O. celebensis , and O. javanicus ) and zebrafish ( Danio rerio ), revealing substantial interspecific variation that broadly corresponded with latitudinal distribution. O. latipes , distributed at the highest latitudes among the tested species, exhibited the greatest cold tolerance, whereas O. celebensis , O. javanicus , and other tropical or low-latitude species showed comparatively low cold tolerance. Our automated system provides a robust, high-throughput platform for thermal tolerance evaluation and, combined with the genetic and genomic resources available in medaka, establishes a foundation for elucidating the molecular mechanisms underlying temperature adaptation in fish.

Janus particles, characterized by their anisotropic structure, exhibit rotational Brownian motion that causes periodic intensity fluctuations (blinking) when suspended in viscous fluids. This phenomenon provides a powerful optical signature for viscosity quantification. In this study, we developed a novel deep learning-based framework to analyze these blinking dynamics for rapid micro-volume viscosity estimation. We first performed extensive simulations of 1-μm Janus particles in fluids with viscosities ranging from 0.1 to 5.0 cP, systematically testing challenging conditions including high translational diffusion, particle aggregation, slow/fast blinking, high noise, and clinical scenarios. Experimental validation was conducted using real glycerol-water mixtures with viscosities ranging between 1.0 and 5.0 cP. A Python-based pipeline combining YOLOv8 instance segmentation and Trackpy tracking accurately detected and tracked individual particles across 140-frame video sequences recorded at 20 frames per second. The Continuous Wavelet Transform (CWT) was then applied to the intensity time-series of each tracked particle to generate time-frequency scalograms (0.05–20 Hz), from which flickering frequencies driven by rotational diffusion were extracted. Finally, a Convolutional Neural Network model, trained on a dataset of 480 CWT-generated intensity heatmaps (20 per viscosity level across 24 viscosity classes), achieved 95.8% accuracy with a mean error of only 0.33% and a 5-fold cross-validation accuracy of 92.7%. Single-blind tests on unseen experimental data yielded viscosity predictions with errors less than 2% with a computational time of less than 2 minutes, with aggregated and non-blinking particles effectively filtered in advance to ensure data quality. This work demonstrates the first seamless integration of instance segmentation, particle tracking, wavelet analysis, and deep learning for automated viscosity prediction, offering a rapid, non-invasive, and robust method with strong potential for applications in microfluidics, biosensing, and clinical diagnostics.

MicroRNAs (miRNAs) are promising biomarkers for the early detection of various diseases, particularly cancer, driving active development of highly sensitive and selective detection technologies. This study aims to establish a novel miRNA detection technique that utilizes image analysis to track the Brownian motion (diffusivity) of fluorescent probe-modified miRNA particles. This method identifies the presence and concentration of miRNAs by exploiting the change in particle size upon hybridization with the target. Furthermore, the use of a probe modified with a photo-crosslinkable artificial nucleic acid (CNV-D) enables the covalent capture of the target miRNA, ensuring high selectivity in biological samples even under stringent washing conditions. By integrating Hybridization Chain Reaction (HCR), the complex size is significantly amplified, dramatically enhancing the detection sensitivity. Consequently, we successfully demonstrated the highly sensitive and specific detection of the cancer biomarker miR-21 in serum, achieving an exceptionally low limit of detection (LOD) of 1 fM. This technology holds great potential to contribute to the early diagnosis of cancer.