Scholarlink

<p>The segmentation of brain tumors, based on hyperspectral imaging (HSI), has been attracting considerable interest for its capability of providing detailed spectral–spatial information about tissue. Current segmentation methods based on HSI have limitations such as spectral noise, illumination changes, high dimensionality of data, and temporal variation between successive surgical frames, which can impact the accuracy of segmentation and its clinical utility. In this work, we suggest a dynamic spectral calibration (DSC) integrated with diffusion-based temporal fusion (DTF) framework to overcome these limitations and achieve interpretable and low latency brain tumor segmentation in HSI. It is based on a spectral-attentive denoising stage using a deep convolutional autoencoder to remove spectral noise while maintaining the most important biochemical and morphological tissue properties. Then, a diffusion-assisted spectral–spatial tumor mapping module is used to iteratively refine the tumor boundary using a reverse diffusion method. A DSC module adaptively recalibrates the spectral responses based on a gradient guided spectral prior optimization to increase robustness in varying illumination conditions during surgery. Moreover, the Hierarchical Spatial–Temporal Integration module integrates multi-scale spatial representations and temporal feature learning via recurrent attention mechanisms and gated recurrent units to ensure temporal coherence between successive frames. The explainable artificial intelligence module is also integrated to offer interpretable visualization of spectral and spatial decision patterns, offering clinical transparency and confidence in the model predictions. Experimental results on hyperspectral brain tumor datasets show that the proposed DSC–DTF framework provides competitive segmentation results with dice similarity coefficient of 97.12%, intersection over union of 94.47%, sensitivity of 96.32%, specificity of 97.18%, Hausdorff distance of 3.42&#xa0;mm, spectral angle mapper of 3.41° and Temporal Consistency Index of 95.25%. The outcomes demonstrate the effectiveness of the proposed framework in enhancing the spectral consistency, segmentation accuracy, and temporal stability of the hyperspectral brain tumor data while maintaining its interpretability and computational efficiency for real-time intraoperative support in analyzing brain tumors.</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

Dynamic Spectral Calibration Integrated with Diffusion Based Temporal Fusion for Accurate and Interpretable Hyperspectral Brain Tumor Segmentation

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Priyanka Gupta,
  • Garima Jaiswal,
  • Ashish Kumar

摘要

The segmentation of brain tumors, based on hyperspectral imaging (HSI), has been attracting considerable interest for its capability of providing detailed spectral–spatial information about tissue. Current segmentation methods based on HSI have limitations such as spectral noise, illumination changes, high dimensionality of data, and temporal variation between successive surgical frames, which can impact the accuracy of segmentation and its clinical utility. In this work, we suggest a dynamic spectral calibration (DSC) integrated with diffusion-based temporal fusion (DTF) framework to overcome these limitations and achieve interpretable and low latency brain tumor segmentation in HSI. It is based on a spectral-attentive denoising stage using a deep convolutional autoencoder to remove spectral noise while maintaining the most important biochemical and morphological tissue properties. Then, a diffusion-assisted spectral–spatial tumor mapping module is used to iteratively refine the tumor boundary using a reverse diffusion method. A DSC module adaptively recalibrates the spectral responses based on a gradient guided spectral prior optimization to increase robustness in varying illumination conditions during surgery. Moreover, the Hierarchical Spatial–Temporal Integration module integrates multi-scale spatial representations and temporal feature learning via recurrent attention mechanisms and gated recurrent units to ensure temporal coherence between successive frames. The explainable artificial intelligence module is also integrated to offer interpretable visualization of spectral and spatial decision patterns, offering clinical transparency and confidence in the model predictions. Experimental results on hyperspectral brain tumor datasets show that the proposed DSC–DTF framework provides competitive segmentation results with dice similarity coefficient of 97.12%, intersection over union of 94.47%, sensitivity of 96.32%, specificity of 97.18%, Hausdorff distance of 3.42 mm, spectral angle mapper of 3.41° and Temporal Consistency Index of 95.25%. The outcomes demonstrate the effectiveness of the proposed framework in enhancing the spectral consistency, segmentation accuracy, and temporal stability of the hyperspectral brain tumor data while maintaining its interpretability and computational efficiency for real-time intraoperative support in analyzing brain tumors.