SKShinya Kojima

Shinya Kojima

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Shinya Kojima has a total of 27 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Evaluation-design effects and reference-CT-defined HU-class errors in brain MRI-to-CT synthesis on SynthRAD2023: a technical note

    by Tatsuya Hayashi, Shinya Kojima, Norio Hayashi

    To quantify how evaluation design affects brain MRI-to-CT synthesis performance, we analyzed 180 MRI/CT pairs from three centers in the SynthRAD2023 Task 1 dataset. A lightweight 2D U-Net was trained with a common validation-based stopping rule under slice-random, patient-random, center-stratified size-matched patient-random, and three center-held-out designs. Comparisons were descriptive. The primary outcome was whole-mask mean absolute error (MAE), supplemented by MAE within reference-CT-defined Hounsfield unit (HU) classes. MAE (mean ± standard deviation) was 94.4 ± 3.0, 100.8 ± 1.3, and 105.2 ± 2.8 HU for the random designs and 187.4 ± 6.0, 160.6 ± 7.1, and 137.4 ± 5.1 HU for centers A, B, and C held out. Slice-random testing yielded lower MAE than patient-random testing. Center-held-out testing yielded higher MAE than the center-stratified comparator, although institutional and acquisition-domain effects could not be separated. Bone errors were high across designs; adipose-range and air errors were elevated in specific center-held-out tests.

  • Open Access

    Large Language Models for the National Radiological Technologist Licensure Examination in Japan: Cross-Sectional Comparative Benchmarking and Evaluation of Model-Generated Items Study

    by Toshimune Ito, Toru Ishibashi, Tatsuya Hayashi, Shinya Kojima, Kazumi Sogabe

    Background: Mock examinations are widely used in health professional education to assess learning and prepare candidates for national licensure. However, instructor-written multiple-choice items can vary in difficulty, coverage, and clarity. Recently, large language models (LLMs) have achieved high accuracy in medical examinations, highlighting their potential for assisting item-bank development; however, their educational quality remains insufficiently characterized. Objective: This study aimed to (1) identify the most accurate LLM for the Japanese National Examination for Radiological Technologists and (2) use the top model to generate blueprint-aligned multiple-choice questions and evaluate their educational quality. Methods: Four LLMs—OpenAI o3, o4-mini, o4-mini-high (OpenAI), and Gemini 2.5 Flash (Google)—were evaluated on all 200 items of the 77th Japanese National Examination for Radiological Technologists in 2025. Accuracy was analyzed for overall items and for 173 nonimage items. The best-performing model (o3) then generated 192 original items across 14 subjects by matching the official blueprint (image-based items were excluded). Subject-matter experts (≥5 y as coordinators and routine mock examination authors) independently rated each generated item on five criteria using a 5-point scale (1=unacceptable, 5=adoptable): item difficulty, factual accuracy, accuracy of content coverage, appropriateness of wording, and instructional usefulness. Cochran Q with Bonferroni-adjusted McNemar tests compared model accuracies, and one-sided Wilcoxon signed-rank tests assessed whether the median ratings exceeded 4. Results: OpenAI o3 achieved the highest accuracy overall (90.0%; 95% CI 85.1%‐93.4%) and on nonimage items (92.5%; 95% CI 87.6%‐95.6%), significantly outperforming o4-mini on the full set ( P =.02). Across models, accuracy differences on the non-image subset were not significant (Cochran Q, P =.10). Using o3, the 192 generated items received high expert ratings for item difficulty (mean, 4.29; 95% CI 4.11‐4.46), factual accuracy (4.18; 95% CI 3.98‐4.38), and content coverage (4.73; 95% CI 4.60‐4.86). Ratings were comparatively lower for appropriateness of wording (3.92; 95% CI 3.73‐4.11) and instructional usefulness (3.60; 95% CI 3.41‐3.80). For these two criteria, the tests did not support a median rating >4 (one-sided Wilcoxon, P =.45 and P ≥.99, respectively). Representative low-rated examples (ratings 1‐2) and the rationale for those scores—such as ambiguous phrasing or generic explanations without linkage to stem cues—are provided in the supplementary materials. Conclusions: OpenAI o3 can generate radiological licensure items that align with national standards in terms of difficulty, factual correctness, and blueprint coverage. However, wording clarity and the pedagogical specificity of explanations were weaker and did not meet an adoptable threshold without further editorial refinement. These findings support a practical workflow in which LLMs draft syllabus-aligned items at scale, while faculty perform targeted edits to ensure clarity and formative feedback. Future studies should evaluate image-inclusive generation, use Application Programming Interface (API)-pinned model snapshots to increase reproducibility, and develop guidance to improve explanation quality for learner remediation.


  • Open Access

    Automatic SNR measurement of brain MR images using a deep learning-based approach

    by Shinya Kojima, Shuntaro Higuchi, Tatsuya Hayashi, Toshiya Kariyasu, Makiko Nishikawa, Hidenori Yamaguchi, Haruhiko Machida

    Background: Signal-to-noise ratio (SNR) is a key metric for evaluating MRI image quality, but conventional measurement methods are time-consuming and operator-dependent. Deep learning offers potential for automating this process. Purpose: To develop and validate a deep learning-based method for automatic SNR measurement from single MRI images. Material and methods: A Pix2Pix framework with a U-Net++ generator and GAN-based discriminator was trained using axial brain MRI images (T1WI, T2WI, and FLAIR) from a 3T scanner. The model generated signal and noise maps from a single image, and SNR maps were computed by pixel-wise division. Whole-brain, white matter (WM), and cerebrospinal fluid (CSF) regions were automatically segmented for regional SNR measurement. The subtraction-map method served as the reference. Structural similarity index (SSIM), correlation coefficients, and Bland–Altman analyses were used to evaluate agreement. Results: Across all sequences, the mean SSIM was 0.95 ± 0.02. SNR values showed strong correlations with the reference method ( r > 0.86) and low relative errors (<7%) for whole-brain, WM, and CSF. Bland–Altman analysis demonstrated a small paired bias and narrow 95% limits of agreement across sequences. Conclusion: The proposed deep learning method enables automatic, accurate, and observer-independent SNR quantification from single MR images, supporting clinical and research image quality evaluation.

  • Open Access

    Large Language Model-Based Generation and Assessment of Radiological Technologist Licensure Exam Items in Japan (Preprint)

    by Toshimune Ito, Toru Ishibashi, Tatsuya Hayashi, Shinya Kojima, Kazumi Sogabe

    BACKGROUND For health professionals, mock examinations are an essential tool for assessing learning outcomes and reinforcing preparation for national licensure. However, traditionally written items are often inconsistent in difficulty, coverage, and clarity. Meanwhile, large language models (LLMs) have demonstrated high accuracy in medical exams and could potentially support the development of item banks, but the educational quality of artificial intelligence (AI)-generated questions remains underexplored. OBJECTIVE To identify the most accurate LLM for the Japanese National Examination for Radiological Technologists and, using that model, to generate and evaluate blueprint‑aligned multiple‑choice questions. METHODS Four LLMs, OpenAI o3, o4-mini, o4-mini-high (OpenAI), and Gemini 2.5 Flash (Google), were evaluated for their accuracy across all 200 items of the 77th Japanese National Examination for Radiological Technologists. The model with the highest accuracy (OpenAI o3) was then used to generate 192 multiple-choice items, adhering to the official blueprint. Expert reviewers rated these AI-generated items across five educational criteria: item difficulty, factual accuracy, content coverage, appropriateness of wording, and instructional usefulness. Statistical analyses were applied to compare model performance and item quality. RESULTS OpenAI o3 achieved the highest accuracy overall (90.0%) and on non-image items (92.5%), significantly outperforming o4-mini on the full set (P = 0.0234). Based on the expert reviewers’ scores, the AI-generated items strongly performed in terms of item difficulty (4.29), factual accuracy (4.18), and content coverage (4.73), whereas significantly lower scores were seen for appropriateness of wording (3.92) and instructional usefulness (3.60) (P < 0.05). CONCLUSIONS OpenAI o3 can generate radiological licensure questions aligned with national standards in terms of difficulty and content accuracy. There are limitations in wording clarity and pedagogical feedback, although these can be addressed through editorial review. This approach can facilitate efficient collaboration between AI and faculty in developing mock exams and also holds promise for supporting scalable, syllabus-aligned assessment in health professional education.

  • Practical signal-to-noise ratio mapping using single clinical MR images

    by Shinya Kojima, Shuntaro Higuchi, Tatsuya Hayashi, Toshiya Kariyasu, Makiko Nishikawa, Hidenori Yamaguchi, Haruhiko Machida

    Accurate signal-to-noise ratio (SNR) measurement is essential for evaluating image quality in magnetic resonance imaging (MRI). While the subtraction-map method provides precise SNR measurements, it requires two consecutive acquisitions, limiting its clinical applicability. This study aims to develop and validate a method for practical SNR measurement using clinical MRI images. The proposed method generates an SNR map by computing a noise-only image from a single MRI image using pixel shifting and edge component removal. The accuracy of our method was compared with the subtraction-map method in three evaluations: (1) optimization of a key parameter for edge component removal, (2) analysis of spatial resolution and SNR level effects, and (3) validation using brain MRI images. The study included brain MRI from 188 patients, and SNR measurements were performed on the resulting images. Correlation coefficients and Bland–Altman analysis were used for comparisons. Parameter optimization identified an optimal threshold for separating noise and edge components. Higher spatial resolution improved accuracy, whereas lower resolution and low SNR conditions led to overestimation. In clinical MRI, the proposed method showed a strong correlation with the subtraction-map method (Spearman r = 0.96), and the highest average error rate was 8.1% in T1-weighted images. Bland–Altman analysis demonstrated good agreement across sequences and regions. This method enables practical SNR estimation from a single image, eliminating the need for repeated acquisitions. While limitations remain in low-SNR or structurally complex regions, the method shows promise as a practical tool for retrospective and routine clinical image quality assessments.


  • Evaluation of deep learning reconstruction on diffusion-weighted imaging quality and apparent diffusion coefficient using an ice-water phantom

    by Tatsuya Hayashi, Shinya Kojima, Toshimune Ito, Norio Hayashi, Hiroshi Kondo, Asako Yamamoto, Hiroshi Oba

    This study assessed the influence of deep learning reconstruction (DLR) on the quality of diffusion-weighted images (DWI) and apparent diffusion coefficient (ADC) using an ice-water phantom. An ice-water phantom with known diffusion properties (true ADC = 1.1 × 10–3 mm2/s at 0 °C) was imaged at various b-values (0, 1000, 2000, and 4000 s/mm2) using a 3 T magnetic resonance imaging scanner with slice thicknesses of 1.5 and 3.0 mm. All DWIs were reconstructed with or without DLR. ADC maps were generated using combinations of b-values 0 and 1000, 0 and 2000, and 0 and 4000 s/mm2. Based on the quantitative imaging biomarker alliance profile, the signal-to-noise ratio (SNRs) in DWIs was calculated, and the accuracy, precision, and within-subject parameter variance (wCV) of the ADCs were evaluated. DLR improved the SNR in DWIs with b-values ranging from 0 to 2000s/mm2; however, its effectiveness was diminished at 4000 s/mm2. There was no noticeable difference in the ADCs of images generated with or without implementing DLR. For a slice thickness of 1.5 mm and combined b-values of 0 and 4000 s/mm2, the ADC values were 0.97 × 10–3and 0.98 × 10–3mm2/s with and without DLR, respectively, both being lower than the true ADC value. Furthermore, DLR enhanced the precision and wCV of the ADC measurements. DLR can enhance the SNR, repeatability, and precision of ADC measurements; however, it does not improve their accuracies.


  • Variability in contrast and apparent diffusion coefficient of kiwifruit used as prostate MRI phantom: 1-week validation

    by Tatsuya Hayashi, Shimpei Yano, Shinya Kojima, Toshimune Ito

    Kiwifruit has been proposed as a phantom for prostate magnetic resonance imaging (MRI) to adjust multiparametric MRI factors. This study evaluated the variability in contrasts and apparent diffusion coefficients (ADCs) via repeated scans of kiwifruits for 1 week. All scans were performed using a 3T MRI system. Six kiwifruits were prepared as phantoms. Each kiwifruit was scanned consecutively for 1 week, and T2-weighted and diffusion-weighted images were acquired. Contrasts between the peripheral placenta (PP) and central placenta (CP), and between the outer pericarp (OP) and CP were calculated. The ADCs of the PP, OP, and CP were also determined. The Friedman test with Scheffé’s post hoc comparison was used to assess contrast and ADC variability over 1 week; no significant differences between the contrasts or ADCs were found. Thus, each kiwifruit phantom exhibited low variability when used as a prostate MRI phantom.


  • Adapting a low-count acquisition of the bone scintigraphy using deep denoising super-resolution convolutional neural network

    by Toshimune Ito, Takafumi Maeno, Hirotatsu Tsuchikame, Masaaki Shishido, Kana Nishi, Shinya Kojima, Tatsuya Hayashi, Kentaro Suzuki


  • Undersampling patterns in k-space for compressed sensing MRI using two-dimensional Cartesian sampling

    by Shinya Kojima, Hiroyuki Shinohara, Takeyuki Hashimoto, Shigeru Suzuki

    In compressed sensing magnetic resonance imaging (CS-MRI), undersampling of k-space is performed to achieve faster imaging. For this process, it is important to acquire data randomly, and an optimal random undersampling pattern is required. However, random undersampling is difficult in two-dimensional (2D) Cartesian sampling. In this study, the effect of random undersampling patterns on image reconstruction was clarified using phantom and in vivo MRI, and a sampling pattern relevant for 2D Cartesian sampling in CS-MRI is suggested. The precision of image restoration was estimated with various acceleration factors and extents for the fully sampled central region of k-space. The root-mean-square error, structural similarity index, and modulation transfer function were measured, and visual assessments were also performed. The undersampling pattern was shown to influence the precision of image restoration, and an optimal undersampling pattern should be used to improve image quality; therefore, we suggest that the ideal undersampling pattern in CS-MRI for 2D Cartesian sampling is one with a high extent for the fully sampled central region of k-space.


  • Iterative image reconstruction that includes a total variation regularization for radial MRI

    by Shinya Kojima, Hiroyuki Shinohara, Takeyuki Hashimoto, Masami Hirata, Eiko Ueno

    This paper presents an iterative image reconstruction method for radial encodings in MRI based on a total variation (TV) regularization. The algebraic reconstruction method combined with total variation regularization (ART_TV) is implemented with a regularization parameter specifying the weight of the TV term in the optimization process. We used numerical simulations of a Shepp–Logan phantom, as well as experimental imaging of a phantom that included a rectangular-wave chart, to evaluate the performance of ART_TV, and to compare it with that of the Fourier transform (FT) method. The trade-off between spatial resolution and signal-to-noise ratio (SNR) was investigated for different values of the regularization parameter by experiments on a phantom and a commercially available MRI system. ART_TV was inferior to the FT with respect to the evaluation of the modulation transfer function (MTF), especially at high frequencies; however, it outperformed the FT with regard to the SNR. In accordance with the results of SNR measurement, visual impression suggested that the image quality of ART_TV was better than that of the FT for reconstruction of a noisy image of a kiwi fruit. In conclusion, ART_TV provides radial MRI with improved image quality for low-SNR data; however, the regularization parameter in ART_TV is a critical factor for obtaining improvement over the FT.