Scholarlink

<p>Cervical cancer remains one of the leading causes of mortality among women worldwide, with early detection playing a critical role in improving therapeutic outcomes. In this study, we propose a novel deep learning framework, CKD-CINet, which integrates multiple components: DeepLabV3+ for spatial feature extraction, Inception-ResNet as the backbone network, a long short-term memory unit for modeling temporal dependencies, an Entangle-Cls module for enhancing feature–spatial interactions, and a multi-scale efficient channel attention module. The model was trained using the focal loss function to address data imbalance and evaluated on publicly available datasets from the IARC Colposcopy Image Bank. The dataset was partitioned into disjoint training, validation, and independent test sets prior to augmentation, and augmentation was applied exclusively to the training subset to prevent data leakage. The proposed CKD-CINet achieved a classification accuracy of 96.3%, a sensitivity of 92.1%, a specificity of 94.6%, and an AUC of 96.7% on an independent test set, demonstrating strong discriminative capability and robustness. Training stability was monitored using an internal validation split with early stopping based on validation loss trends to mitigate overfitting. Ablation studies further validated the contribution of each architectural component to the overall performance. Comparative evaluation was conducted under an identical dataset and preprocessing protocol to ensure fair benchmarking. The integration of spatio-temporal learning with attention mechanisms significantly improved the model’s ability to accurately detect high-grade lesions. From a computational perspective, the proposed CKD-CINet integrates multiple deep learning components, including segmentation, classification, attention mechanisms, and temporal modeling. Training such cascaded architectures on high-resolution colposcopy images involves substantial computational workload and benefits from GPU-accelerated parallel processing. This computational strategy enables efficient model optimization and supports near real-time inference for large-scale cervical cancer screening applications.</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

Improved detection of abnormal cervical tissue growth using cascade knowledge distillation based deep learning

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Alireza Taremi,
  • Mohammad Reza Ebrahimi Dishabi,
  • Hamed Pezeshki,
  • Mojgan Karimi Zarchi

摘要

Cervical cancer remains one of the leading causes of mortality among women worldwide, with early detection playing a critical role in improving therapeutic outcomes. In this study, we propose a novel deep learning framework, CKD-CINet, which integrates multiple components: DeepLabV3+ for spatial feature extraction, Inception-ResNet as the backbone network, a long short-term memory unit for modeling temporal dependencies, an Entangle-Cls module for enhancing feature–spatial interactions, and a multi-scale efficient channel attention module. The model was trained using the focal loss function to address data imbalance and evaluated on publicly available datasets from the IARC Colposcopy Image Bank. The dataset was partitioned into disjoint training, validation, and independent test sets prior to augmentation, and augmentation was applied exclusively to the training subset to prevent data leakage. The proposed CKD-CINet achieved a classification accuracy of 96.3%, a sensitivity of 92.1%, a specificity of 94.6%, and an AUC of 96.7% on an independent test set, demonstrating strong discriminative capability and robustness. Training stability was monitored using an internal validation split with early stopping based on validation loss trends to mitigate overfitting. Ablation studies further validated the contribution of each architectural component to the overall performance. Comparative evaluation was conducted under an identical dataset and preprocessing protocol to ensure fair benchmarking. The integration of spatio-temporal learning with attention mechanisms significantly improved the model’s ability to accurately detect high-grade lesions. From a computational perspective, the proposed CKD-CINet integrates multiple deep learning components, including segmentation, classification, attention mechanisms, and temporal modeling. Training such cascaded architectures on high-resolution colposcopy images involves substantial computational workload and benefits from GPU-accelerated parallel processing. This computational strategy enables efficient model optimization and supports near real-time inference for large-scale cervical cancer screening applications.