Faculty Profiles - ODA, Masahiro

Spiking MU-Net: Toward Low-Power and Efficient Microscopic Image Segmentation Reviewed

Wang L., Wu Y., Oda M., Hayashi Y., Mori K.

Lecture Notes in Computer Science Vol. 16318 LNCS page： 163 - 173 2026

Language：English Publishing type：Research paper (international conference proceedings) Publisher：Lecture Notes in Computer Science

In contemporary biomedical research, microscopic image segmentation has emerged as an indispensable analytical task. However, complex morphologies and diverse sizes of cellular structures make accurate cell segmentation challenging. Although deep learning based methods have shown great potential in microscopic image segmentation, they come with high computational complexity, reliance on extensive GPU resources, large parameter counts, and high power consumption. To address these challenges, this paper introduces a novel, computationally efficient framework that integrates the Mamba with Spiking Neural Networks (SNNs) for biomedical applications. Spiking MU-Net maintains accuracy comparable to state-of-the-art methods while significantly reducing computational load, parameter size, and power consumption. Specifically, (1) We propose a spiking neuron-based model that integrates Mamba and convolutional modules, enabling discrete, addition-only operations. This design significantly reduces computational overhead and parameter count, paving the way for real-time microscopy analysis. (2) Experiments on multiple cell datasets demonstrated that our method achieved high accuracy with a DSC of 83% while significantly reducing energy consumption to just 3.22mJ. (3) Our approach attained accuracy comparable to ANN-based networks, with a DSC improvement of 1.04%, while achieving the lowest energy consumption among SNN-based segmentation models, reducing parameters by 10.21M and energy by 0.29mJ.

DOI： 10.1007/978-3-032-13961-0_17

Fully automated segmentation of substantia nigra toward longitudinal analysis of Parkinson's disease

YOLOv5 Attention Analysis for Anterior Eye Disease Classification: Grad-CAM++ Feature Importance and Cut-and-Paste Validation Reviewed

Yoshiyuki Kitaguchi, Yuta Ueno, Takefumi Yamaguchi, Hiroki Maehara, Dai Miyazaki, Ryohei Nejima, Takenori Inomata, Naoko Kato, Tai-ichiro Chikama, Jun Ominato, Tatsuya Yunoki, Kinya Tsubota, Masahiro Oda, Kensaku Mori, Yu Yoshinaga, Rikako Iwasaki, Kohji Nishida, Tetsuro Oshika

Journal of Imaging Informatics in Medicine 2025.10

Deep Learning Model for Extensive Diagnosis of Corneal Deposits. Reviewed

Taki Y, Ueno Y, Oda M, Kitaguchi Y, Maehara H, Yamaguchi T

Cornea 2025.10

Generative AI and foundation models in medical image Invited Reviewed Open Access

Masahiro Oda

Radiological Physics and Technology Vol. 18 page： 937 - 948 2025.10

Hu, T; Itoh, H; Oda, M; Saiki, S; Kamagata, K; Ishikawa, K; Sako, W; Hattori, N; Aoki, S; Mori, K

INTERNATIONAL JOURNAL OF COMPUTER ASSISTED RADIOLOGY AND SURGERY 2025.10

Language：English Publisher：International Journal of Computer Assisted Radiology and Surgery

Purpose: A fully automated segmentation of substantia nigra (SN) is an essential task for the development of an explainable computer-aided diagnosis system of Parkinson’s disease (PD). Since anatomical alterations of SN are vital information in PD diagnosis, a precise segmentation model should have generalization ability against spatiotemporal changes. To satisfy these requirements, we propose a fully automated pipeline with several new techniques for a volumetric image obtained by neuromelanin magnetic resonance imaging. Methods: We develop a pipeline by integrating SN-prior probability estimation into the decision of the SN-contained region of interest. An estimated SN-prior probability is further fed into a new priority attention mechanism as a gating signal in our segmentation model. Furthermore, we introduce test-time dropout to improve a segmentation model’s accuracy and generalization ability. To evaluate the model’s generalization ability, we collected principal and external datasets with longitudinal scans of the same PD patients. Results: Our segmentation model achieved averaged Dice scores of 0.845 and 0.851 for SN hyperintense regions in the principal and external datasets, respectively. These results demonstrated the best generalization ability in our comparative evaluations. Thresholding the number of voxels in the SN hyperintense regions, we also evaluated the segmentation results in automated PD identification. The PD identification achieved the areas under the receiver operating characteristic curves of 0.755 and 0.726 by our pipeline’s output and the ground truth, respectively. Conclusions: The proposed pipeline, where we integrated SN-prior probability estimation, priority attention mechanism and test-time dropout to our segmentation model, achieved accurate SN segmentation with high generalization ability for our longitudinal data: the principal and external datasets. As demonstrated in the validation with the automated PD identification, our pipeline has the potential for improving the performance of PD diagnosis via further large-scale longitudinal analysis.

DOI： 10.1007/s11548-025-03451-9

Long-term impact of computer-aided adenoma detection: a prospective observational study Reviewed

Okumura, T; Kudo, S; Ide, Y; Kato, S; Miyata, Y; Takisima, K; Takashina, Y; Minegishi, Y; Abe, M; Sakurai, T; Koyama, Y; Kato, K; Maeda, Y; Ogawa, Y; Ichimasa, K; Ogata, N; Hayashi, T; Wakamura, K; Baba, T; Nemoto, T; Miyachi, H; Itoh, H; Oda, M; Mori, K; Misawa, M

ENDOSCOPY 2025.9

Language：English Publishing type：Research paper (scientific journal) Publisher：Endoscopy

Background Computer-aided detection (CADe) systems have improved the adenoma detection rate (ADR); however, concerns about its long-term effect on endoscopists’ performance without CADe and potential deskilling remain unaddressed. This study evaluated the impact of CADe on the learning curve of endoscopists for adenoma detection. Methods This propensity score-matching, prospective, single-center, observational study was conducted from January 2021 to December 2023. CADe systems were installed in half of the endoscopy units, and patients were equally distributed between the rooms. Patients aged ≥20 years scheduled for colonoscopy were included, excluding those with polyposis, inflammatory bowel disease, known polyps, incomplete colonoscopy, emergency cases, or previous colorectal surgery, and those examined by novices. Endoscopists were classified as high detectors (ADR ≥25%) or low detectors (ADR <25%) based on the ADR recorded before CADe implementation. To assess skill acquisition and transfer, the primary outcome was the change in ADR over time, as measured by cumulative summation (CUSUM) analysis, in both CADe and non-CADe procedures. Results Of 18962 patients who underwent colonoscopy, 13245 patients were excluded, and of the 5717 patients initially enrolled, 4712 (CADe group, n =2356; non-CADe group, n =2356) were analyzed after propensity score matching. CUSUM analysis showed that both high and low detectors achieved enhanced detection performance for CADe procedures. Among non-CADe procedures, high detectors had accelerated learning curves, indicating they maintained a higher ADR, whereas low detectors showed no significant change in their learning trajectory. Conclusions After CADe implementation, the detection rate in procedures performed without CADe was maintained and did not decline over time.

DOI： 10.1055/a-2661-2624

Synthetic data generation with Worley-Perlin diffusion for robust subarachnoid hemorrhage detection in imbalanced CT Datasets Reviewed

Lu, ZY; Hu, T; Oda, M; Fuse, Y; Saito, R; Jinzaki, M; Mori, K

INTERNATIONAL JOURNAL OF COMPUTER ASSISTED RADIOLOGY AND SURGERY 2025.9

Language：English Publishing type：Research paper (scientific journal) Publisher：International Journal of Computer Assisted Radiology and Surgery

Purpose: In this paper, we propose a novel generative model to produce high-quality SAH samples, enhancing SAH CT detection performance in imbalanced datasets. Previous methods, such as cost-sensitive learning and previous diffusion models, suffer from overfitting or noise-induced distortion, limiting their effectiveness. Accurate SAH sample generation is crucial for better detection. Methods: We propose the Worley–Perlin Diffusion Model (WPDM), leveraging Worley–Perlin noise to synthesize diverse, high-quality SAH images. WPDM addresses limitations of Gaussian noise (homogeneity) and Simplex noise (distortion), enhancing robustness for generating SAH images. Additionally, WPDMFast optimizes generation speed without compromising quality. Results: WPDM effectively improved classification accuracy in datasets with varying imbalance ratios. Notably, a classifier trained with WPDM-generated samples achieved an F1-score of 0.857 on a 1:36 imbalance ratio, surpassing the state of the art by 2.3 percentage points. Conclusion: WPDM overcomes the limitations of Gaussian and Simplex noise-based models, generating high-quality, realistic SAH images. It significantly enhances classification performance in imbalanced settings, providing a robust solution for SAH CT detection.

DOI： 10.1007/s11548-025-03482-2

Enhancing YOLO for laparoscopic tool detection: novel data augmentation and structural modifications addressing mis-detection of bifurcated targets

GPU-accelerated deformation mapping in hybrid organ models for real-time simulation Reviewed Open Access

Rintaro Miyazaki, Yuichiro Hayashi, Masahiro Oda, Kensaku Mori

International Journal of Computer Assisted Radiology and Surgery 2025.7

Liu, YZ; Hayashi, Y; Oda, M; Mori, K

INTERNATIONAL JOURNAL OF COMPUTER ASSISTED RADIOLOGY AND SURGERY Vol. 20 ( 9 ) page： 1899 - 1910 2025.7

Language：English Publisher：International Journal of Computer Assisted Radiology and Surgery

Purpose : Laparoscopic tool detection is vital for assistance of minimally invasive surgeries, aiding tasks like tool pose estimation and surgical navigation. This study enhances YOLO models for better detection of bifurcated targets (BT) in such procedures, addressing the issue of mis-detection of bifurcated targets (MDBT) where BT tips are misidentified as separate entities or overlooked. Methods : We proposed a data augmentation strategy, Random Target Masking, to prevent the model from identifying BT tips as separate laparoscopic tools. Mixup Plus was developed to balance instance count across categories with varying BT proportions. Additionally, we employed the Space-to-Depth Convolution block for downsampling to curb the model’s tendency to overlook small-sized BT tips. Results : The YOLOv8 model featuring our modifications, tested on our dataset derived from EndoVis17 and EndoVis18, showed improvement in both mAP50 and mAP50:95 metrics on the test dataset. On the BT test dataset specifically, mAP50 and mAP50:95 metrics improved by nearly 0.2 and 0.1, respectively. For the Clip Applier category, which has the fewest instances (fewer than 100 instances in the entire training and test dataset), the YOLOv8n model incorporating our proposed modifications increased AP50 from 0.0251 to 0.457. Conclusion : This study focused on improving BT detection accuracy in laparoscopic tool detection using YOLO models, incorporating RTM and MUP data augmentation techniques along with SPD-Conv block integration. Experimental evaluations based on the EndoVis datasets validated the enhancements. The ablation study confirmed the effectiveness of each proposed improvement, particularly highlighting the distinct advantages of the proposed data augmentation methods.

DOI： 10.1007/s11548-025-03352-x

Diffusion-driven distillation and contrastive learning for class-incremental semantic segmentation of laparoscopic images

Zhao, XK; Hayashi, Y; Oda, M; Kitasaka, T; Mori, K

INTERNATIONAL JOURNAL OF COMPUTER ASSISTED RADIOLOGY AND SURGERY Vol. 20 ( 7 ) page： 1551 - 1560 2025.7

Language：English Publisher：International Journal of Computer Assisted Radiology and Surgery

Purpose: Understanding anatomical structures in laparoscopic images is crucial for various types of laparoscopic surgery. However, creating specialized datasets for each type is both inefficient and challenging. This highlights the clinical significance of exploring class-incremental semantic segmentation (CISS) for laparoscopic images. Although CISS has been widely studied in diverse image datasets, in clinical settings, incremental data typically consists of new patient images rather than reusing previous images, necessitating a novel algorithm. Methods: We introduce a distillation approach driven by a diffusion model for CISS of laparoscopic images. Specifically, an unconditional diffusion model is trained to generate synthetic laparoscopic images, which are then incorporated into subsequent training steps. A distillation network is employed to extract and transfer knowledge from networks trained in earlier steps. Additionally, to address the challenge posed by the limited semantic information available in individual laparoscopic images, we employ cross-image contrastive learning, enhancing the model’s ability to distinguish subtle variations across images. Results: Our method was trained and evaluated on all 11 anatomical structures from the Dresden Surgical Anatomy Dataset, which presents significant challenges due to its dispersed annotations. Extensive experiments demonstrate that our approach outperforms other methods, especially in difficult categories such as the ureter and vesicular glands, where it surpasses even supervised offline learning. Conclusion: This study is the first to address class-incremental semantic segmentation for laparoscopic images, significantly improving the adaptability of segmentation models to new anatomical classes in surgical procedures.

DOI： 10.1007/s11548-025-03405-1