Scholarlink

<p>Medical imaging utilizes multiple imaging modalities as it offers diverse and complementary information about tissues, organs, or tumours. Combining this multi-information to enhance segmentation is known as multimodal image segmentation. This task is done by processing images from different imaging devices and it is an important procedure in medical image analysis. A key element in image analysis is image registration and its objective is to align the coordinate system of one image with another. For multimodal medical images, effective registration is vital for image-guided surgery, treatment planning and clinical diagnosis, as it fuses the information from different image modalities. However, due to the inherent differences in image characteristics from different modalities, finding an accurate method for matching these images remains a difficult problem. Many researchers have explored the use of deep learning methods, especially in medical image segmentation and registration. Despite the progress, many existing solutions are limited to segment specific anatomical regions and are restricted by their generalizability across multiple imaging modalities and they often require significant computational resources, particularly in clinical applications. Therefore, a novel joint adaptive framework is introduced for performing registration and segmentation using multimodal images. At first, the available benchmark sources, the multimodal images are aggregated. Further, the images are forwarded to the Adaptive 3D Transformer-based Residual Registration with Segmentation Network (ATRSNet) for performing the image registration and segmentation. Here, a Hybridized Kookaburra with Lotus Effect Optimization Algorithm (HK-LEOA) is employed to optimally determine the parameters of the A3D-TRSNet technique. This enhances the overall registration and segmentation performance rates. At last, the experiments are performed to ensure the effectualness of the presented work. The statistical outcomes of the developed model outperform 4.27%, 4.13%, 3.51% and 2.32% better performance than ZOA-ATRSNet, WOA-ATRSNet, KOA-ATRSNet and LEOA-ATRSNet, respectively in terms of best measure. Additionally, the dice score of the developed model achieves 93.57 ± 0.49, which increases the accuracy rate in the multi-modal image registration framework.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel joint adaptive network for multimodal image registration and segmentation using 3D transformer-based residual registration with segmentation network

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Vilas Vasudeo Deotare,
  • Ujwala Sushil Ghodeswar,
  • Alampally Sreedevi

摘要

Medical imaging utilizes multiple imaging modalities as it offers diverse and complementary information about tissues, organs, or tumours. Combining this multi-information to enhance segmentation is known as multimodal image segmentation. This task is done by processing images from different imaging devices and it is an important procedure in medical image analysis. A key element in image analysis is image registration and its objective is to align the coordinate system of one image with another. For multimodal medical images, effective registration is vital for image-guided surgery, treatment planning and clinical diagnosis, as it fuses the information from different image modalities. However, due to the inherent differences in image characteristics from different modalities, finding an accurate method for matching these images remains a difficult problem. Many researchers have explored the use of deep learning methods, especially in medical image segmentation and registration. Despite the progress, many existing solutions are limited to segment specific anatomical regions and are restricted by their generalizability across multiple imaging modalities and they often require significant computational resources, particularly in clinical applications. Therefore, a novel joint adaptive framework is introduced for performing registration and segmentation using multimodal images. At first, the available benchmark sources, the multimodal images are aggregated. Further, the images are forwarded to the Adaptive 3D Transformer-based Residual Registration with Segmentation Network (ATRSNet) for performing the image registration and segmentation. Here, a Hybridized Kookaburra with Lotus Effect Optimization Algorithm (HK-LEOA) is employed to optimally determine the parameters of the A3D-TRSNet technique. This enhances the overall registration and segmentation performance rates. At last, the experiments are performed to ensure the effectualness of the presented work. The statistical outcomes of the developed model outperform 4.27%, 4.13%, 3.51% and 2.32% better performance than ZOA-ATRSNet, WOA-ATRSNet, KOA-ATRSNet and LEOA-ATRSNet, respectively in terms of best measure. Additionally, the dice score of the developed model achieves 93.57 ± 0.49, which increases the accuracy rate in the multi-modal image registration framework.