Radiology report generation is an important application of artificial intelligence (AI), as the interpretation of medical images and the production of clinically relevant reports are time-consuming and cognitively demanding tasks, especially in high-volume settings such as chest X-ray screening. Recent advances in large language models (LLMs) and vision-language models (VLMs) have enabled substantial progress in automated medical report generation. However, most existing medical foundation models are trained primarily on English datasets, limiting their practicality in non-English-speaking regions such as Japan. Publicly available radiology datasets are overwhelmingly English, while constructing large-scale non-English datasets is costly and difficult because of translation effort and medical data privacy constraints. Existing adaptation methods, such as fine-tuning and continued pre-training, typically require large amounts of in-domain data. This makes them difficult to apply in low-resource medical language settings where large-scale annotated datasets are unavailable. Accordingly, this study asks: how can an accurate Japanese chest X-ray radiology report generator be developed without access to large-scale, curated Japanese medical image-report pairs? To address this challenge, we propose JRadiEvo, a Japanese chest X-ray report generation model built through evolutionary optimization of model merging. By combining pretrained models with complementary strengths in vision-language alignment, medical knowledge, and Japanese generation, JRadiEvo enables data-efficient adaptation without large-scale training. To the best of our knowledge, this is the first attempt to build a non-English medical vision-language model through evolutionary optimization of model merging. Despite using only 50 translated training samples from publicly available data, JRadiEvo outperforms CheXagent, a state-of-the-art model trained on approximately 8.5 million samples, in ROUGE-L and METEOR metrics. These results provide a proof of concept for extreme data-efficient adaptation in low-resource medical languages.

An overview is provided of on-going production of regional atmospheric reanalysis data RRJ-ClimCORE for Japan and its surrounding maritime domain. The data can reproduce the past atmospheric conditions four-dimensionally with high spatial and temporal resolutions, through assimilating fine-resolution observational data into a state-of-the-art operational regional forecast system. The reanalysis data will have wide-ranging potential applications in society, including disaster prevention and risk management planning, and system designing for renewable energy through deep learning.

To account for the unique characteristics and limitations of the human visual system (HVS) when perceiving images, a variety of perceptual quality metrics have been proposed in the literature. Tailoring rate-distortion (RD) optimization for each metric is cumbersome and time-consuming. In this paper, we propose a general RD-optimization strategy called “transform domain bounding box” (BB) that can easily adapt to different quality metrics for JPEG-like block-based encoding of images. First, we define an objective function that is a weighted sum of the l0-norm of the transform coefficients (a proxy for rate) and distortion from the transform domain representation. Next, for a given distortion target τ, we define a don't care region (DCR) that specifies a search region of representations with distortion ≤τ. We then show that the sparsest transform domain representation (lowest encoding rate) inside a BB that tightly contains the DCR can be constructed efficiently. Varying τ to induce different DCRs and corresponding BBs results in a set of constructed sparse representations of different sparsity counts, and the one that optimally trades off rate and distortion can be easily identified as solution to our objective. We show that our proposed BB strategy can be easily re-targeted for three common quality metrics: MSE, MSE-HVS-M and SSIM. Experimental results show that our BB strategy outperformed unoptimized JPEG compression by up to 1dB in PSNR when distortion metric is MSE, up to 2dB when metric is MSE-HVS-M, and up to 0.005 when metric is SSIM.

Compression of depth maps is important for “texture plus depth” format of multiview images, which enables synthesis of novel intermediate views via depth-image-based rendering (DIBR) at decoder. Previous depth map coding schemes exploit unique depth data characteristics to compactly and faithfully reproduce the original signal. In contrast, since depth map is only a means to the end of view synthesis and not itself viewed, in this paper we explicitly manipulate depth values, without causing severe synthesized view distortion, in order to maximize representation sparsity in the transform domain for compression gain - we call this process transform domain spar-sification (TDS). Specifically, for each pixel in the depth map, we first define a quadratic penalty function, with minimum at ground truth depth value, based on synthesized view's distortion sensitivity to the pixel's depth value during DIBR. We then define an objective for a depth signal in a block as a weighted sum of: i) signal's sparsity in the transform domain, and ii) per-pixel synthesized view distortion penalties for the chosen signal. Given that sparsity (l0-norm) is non-convex and difficult to optimize, we replace the l0-norm in the objective with a computationally inexpensive weighted l2-norm; the optimization is then an unconstrained quadratic program, solvable via a set of linear equations. For the weighted l2-norm to promote sparsity, we solve the optimization iteratively, where at each iteration weights are readjusted to mimic sparsity-promoting lτ-norm, 0 ≤ τ ≤ 1. Using JPEG as an example transform codec, we show that our TDS approach gained up to 1.7dB in rate-distortion performance for the interpolated view over compression of unaltered depth maps.

Depth map compression is important for compact “texture-plus-depth” representation of a 3D scene, where texture and depth maps captured from multiple camera viewpoints are coded into the same format. Having received such format, the decoder can synthesize any novel intermediate view using texture and depth maps of two neighboring captured views via depth-image-based rendering (DIBR). In this paper, we combine two previously proposed depth map compression techniques that promote sparsity in the transform domain for coding gain-graph-based transform (GBT) and transform domain sparsification (TDS) - together under one unified optimization framework. The key to combining GBT and TDS is to adaptively select the simplest transform per block that leads to a sparse representation. For blocks without detected prominent edges, the synthesized view's distortion sensitivity to depth map errors is low, and TDS can effectively identify a sparse depth signal in fixed DCT domain within a large search space of good signals with small synthesized view distortion. For blocks with detected prominent edges, the synthesized view's distortion sensitivity to depth map errors is high, and the search space of good depth signals for TDS to find sparse representations in DCT domain is small. In this case, GBT is first performed on a graph defining all detected edges, so that filtering across edges is avoided, resulting in a sparsity count ρ in GBT. We then incrementally add the most important edge to an initial no-edge graph, each time performing TDS in the resulting GBT domain, until the same sparsity count ρ is achieved. Experimentation on two sets of multiview images showed gain of up to 0.7dB in PSNR in synthesized view quality compared to previous techniques that employ either GBT or TDS alone.

Regional weather forecasting demands fast simulation over fine-grained grids, resulting in extremely memory- bottlenecked computation, a difficult problem on conventional supercomputers. Early work on accelerating mainstream weather code WRF using GPUs with their high memory performance, however, resulted in only minor speedup due to partial GPU porting of the huge code. Our full CUDA porting of the high- resolution weather prediction model ASUCA is the first such one we know to date; ASUCA is a next-generation, production weather code developed by the Japan Meteorological Agency, similar to WRF in the underlying physics (non-hydrostatic model). Benchmark on the 528 (NVIDIA GT200 Tesla) GPU TSUBAME Supercomputer at the Tokyo Institute of Technology demonstrated over 80-fold speedup and good weak scaling achieving 15.0 TFlops in single precision for 6956 x 6052 x 48 mesh. Further benchmarks on TSUBAME 2.0, which will embody over 4000 NVIDIA Fermi GPUs and deployed in October 2010, will be presented.