Scholarlink

<p>Pulmonary nodules (PNs) are associated with lung cancer and has high mortality rate, mostly due to late-stage diagnosis and complex mechanism for visual characteristics in computed tomography (CT) for automated classification. Its accurate early detection is limited due to problems like insufficient annotated datasets, visual similarity of benign and malignant nodules and poor generalization in traditional deep learning models. In response to these limitations, the study introduces a novel framework called BCM-FSS-MHA (Base-Class Mining–Few-Shot Segmentation–Multi-Head Attention) where base-class knowledge retention learning, adaptive FSS and attention based feature refinement are integrated for segmentation and classification. MHA mechanism is adapted to the refined feature maps for learning discriminative representations of nodules from texture information, shape information and boundary appearance. Using the publicly accessible Lung Nodule Analysis (LUNA16) dataset, our proposed method accomplishes efficient CT scan pre-processing, generalized few-shot semantic segmentation and attention-driven feature aggregation that extends XGBoost to perform classification. This model is able to achieve a Segmentation IoU (Intersection over union) of 80.3%, DSC (Dice similarity coefficient) of 90.6%, Classification accuracy of 98.57% and an AUC of 99.1% thereby surpassing existing baseline models.</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

BCM-FSS-MHA: attention-enhanced few-shot learning approach for solitary pulmonary nodule detection in low-dose CT imaging

  • 期刊论文

  • 发表日期：

    2025年12月10日

查看全文
  • Gurbhej Singh,
  • Geetika Sharma,
  • Anupam Mittal

摘要

Pulmonary nodules (PNs) are associated with lung cancer and has high mortality rate, mostly due to late-stage diagnosis and complex mechanism for visual characteristics in computed tomography (CT) for automated classification. Its accurate early detection is limited due to problems like insufficient annotated datasets, visual similarity of benign and malignant nodules and poor generalization in traditional deep learning models. In response to these limitations, the study introduces a novel framework called BCM-FSS-MHA (Base-Class Mining–Few-Shot Segmentation–Multi-Head Attention) where base-class knowledge retention learning, adaptive FSS and attention based feature refinement are integrated for segmentation and classification. MHA mechanism is adapted to the refined feature maps for learning discriminative representations of nodules from texture information, shape information and boundary appearance. Using the publicly accessible Lung Nodule Analysis (LUNA16) dataset, our proposed method accomplishes efficient CT scan pre-processing, generalized few-shot semantic segmentation and attention-driven feature aggregation that extends XGBoost to perform classification. This model is able to achieve a Segmentation IoU (Intersection over union) of 80.3%, DSC (Dice similarity coefficient) of 90.6%, Classification accuracy of 98.57% and an AUC of 99.1% thereby surpassing existing baseline models.