Scholarlink

<p>This work introduces a hybrid attention-driven deep learning model for real-time scenarios within a resource-constrained environment in healthcare. In image-based healthcare applications, i.e., medical image classification, segmentation, and abnormality identification, the model is supposed to decrease the computational costs and memory footprint without compromising or, to the contrary, enhancing predictive quality. The proposed model outlines five important elements such as a cascaded scheme that facilitates the early exit decision making on diagnosis cases that are of a diagnostically simple nature and makes use of the Low-Rank Tensor Factorisation scheme to remove the spatial and channel-wise redundancy of the feature map. Quantised intermediate buffers are also applied to minimise intermediate activation; the activations are quantised at eight bits, which minimises the memory needs at inference. In addition to this, depthwise separable convolutions are also employed for better use of the computational resources. Due to the experimental studies of the medical data, including dermoscopic and chest X-ray images, one can confirm that the proposed model offers a decent combination of diagnostic performance, inference latency, and resource utilisation. The characteristics make it particularly suitable to be used in boundary-based clinical environments where effective processing and rapid response are the most important.</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

LRA-CNN: a novel lite rank attention based CNN architecture for brain tumour classification

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Abhisek Hazra,
  • Sitanath Biswas,
  • Shubhashree Sahoo

摘要

This work introduces a hybrid attention-driven deep learning model for real-time scenarios within a resource-constrained environment in healthcare. In image-based healthcare applications, i.e., medical image classification, segmentation, and abnormality identification, the model is supposed to decrease the computational costs and memory footprint without compromising or, to the contrary, enhancing predictive quality. The proposed model outlines five important elements such as a cascaded scheme that facilitates the early exit decision making on diagnosis cases that are of a diagnostically simple nature and makes use of the Low-Rank Tensor Factorisation scheme to remove the spatial and channel-wise redundancy of the feature map. Quantised intermediate buffers are also applied to minimise intermediate activation; the activations are quantised at eight bits, which minimises the memory needs at inference. In addition to this, depthwise separable convolutions are also employed for better use of the computational resources. Due to the experimental studies of the medical data, including dermoscopic and chest X-ray images, one can confirm that the proposed model offers a decent combination of diagnostic performance, inference latency, and resource utilisation. The characteristics make it particularly suitable to be used in boundary-based clinical environments where effective processing and rapid response are the most important.