Scholarlink

<p>Histopathological imaging, a crucial component of diagnostic medical imaging, faces significant challenges in low-resource settings due to a shortage of annotated datasets and the computational inefficiency of conventional deep learning models such as convolutional neural networks (CNNs). To address these limitations, we propose a Vision Transformer (ViT)-based framework that achieves high diagnostic accuracy with minimal data while ensuring computational efficiency suitable for underdeveloped healthcare infrastructures. The framework incorporates self-attention mechanisms to capture global histopathological patterns, cross-domain pretraining to reduce dependence on labeled data, synthetic histology generation using generative adversarial networks (GANs) to improve robustness against imaging artifacts, and hardware-aware model compression (quantization and pruning) to minimize GPU constraints. Performance was validated on biopsy images with simulated artifacts (80% masking and noise injection), where the model achieved 98% classification accuracy (95% CI: 96.5–99.2%) using only 500 annotated examples, surpassing CNNs (89%, p&#xa0;&lt;&#xa0;0.01) and pathologists (92%, p&#xa0;&lt;&#xa0;0.05). The framework also reduced training time by 40% (8.2 vs. 13.7 h) and GPU memory usage by 35% (4.8 vs. 7.4 GB), while synthetic data integration decreased variance in accuracy by 60% (SD: ±&#xa0;1.2% vs. ±&#xa0;3.1%). These results demonstrate that the proposed model not only outperforms existing methods in terms of accuracy and computational efficiency but also provides a scalable and equitable solution for cancer diagnostics in resource-limited settings, thereby addressing critical disparities in global oncology.</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

Resource-Efficient Cancer Diagnosis in a Resource-Limited Context: Vision Transformer-Based Framework of Data-scarce Histopathological Image Classification with Deep Neural Network

  • 期刊论文

  • 发表日期：

    2026年1月23日

查看全文
  • Namrata Singh,
  • Meenakshi Srivastava,
  • Geetika Srivastava

摘要

Histopathological imaging, a crucial component of diagnostic medical imaging, faces significant challenges in low-resource settings due to a shortage of annotated datasets and the computational inefficiency of conventional deep learning models such as convolutional neural networks (CNNs). To address these limitations, we propose a Vision Transformer (ViT)-based framework that achieves high diagnostic accuracy with minimal data while ensuring computational efficiency suitable for underdeveloped healthcare infrastructures. The framework incorporates self-attention mechanisms to capture global histopathological patterns, cross-domain pretraining to reduce dependence on labeled data, synthetic histology generation using generative adversarial networks (GANs) to improve robustness against imaging artifacts, and hardware-aware model compression (quantization and pruning) to minimize GPU constraints. Performance was validated on biopsy images with simulated artifacts (80% masking and noise injection), where the model achieved 98% classification accuracy (95% CI: 96.5–99.2%) using only 500 annotated examples, surpassing CNNs (89%, p < 0.01) and pathologists (92%, p < 0.05). The framework also reduced training time by 40% (8.2 vs. 13.7 h) and GPU memory usage by 35% (4.8 vs. 7.4 GB), while synthetic data integration decreased variance in accuracy by 60% (SD: ± 1.2% vs. ± 3.1%). These results demonstrate that the proposed model not only outperforms existing methods in terms of accuracy and computational efficiency but also provides a scalable and equitable solution for cancer diagnostics in resource-limited settings, thereby addressing critical disparities in global oncology.