Scholarlink

Background <p>Digital breast tomosynthesis (DBT) offers a better ability to visualise lesions than conventional mammography and is becoming a valuable imaging modality for detecting breast cancers. Although DBT reduces tissue overlap compared to conventional 2D mammography, automated diagnosis remains challenging due to volumetric complexity and variability in lesion appearance.</p> Objective <p>We present a novel approach, Three-dimensional Multi-scale Features Fusion (3DMSFF), for classifying DBT volumes into benign and malignant classes according to a fusion of deep learning models with a new multi-scale features fusion strategy. The new strategy combines two deep learning models to extract global context knowledge and localised fine-grained features essential for precise tumour classification.</p> Methodology <p>The proposed model consists of a two-stream 3D deep learning model to learn complementary information from 3D volumes of DBT; (i) Local Feature Extraction - Convolutional Neural Networks (LFE-CNN) is a Convolutional Neural Network (CNN) branch that extracts low-level information like textural and structural information with residual connections, (ii) Global Feature Extraction – Convolutional Neural Networks (GFE-CNN) uses separable convolutions and squeeze and excitation (SE) blocks in depth to highlight large-scale discriminative features and contextual correlations within 3D breast volume. To reduce redundancy and computational load, each DBT volume is first processed through entropy-based ranking to extract the 20 most informative slices, which are then stacked to form a refined 3D subvolume used as model input. The features learnt from the two branches are combined through a feature concatenation technique, and fully connected layers are employed to make the final classification. The model is trained with focal loss to address the class imbalance issue.</p> Results <p>The 3DMSFF architecture is evaluated on a dataset of 3D DBT volumes labelled benign or malignant. The results present a significant improvement in classification performance compared to the standard 3D CNN architecture, as testified by the increased accuracy, sensitivity, and Area Under the Receiver Operating Characteristic curve (AUC-ROC) performance. The proposed work 3DMSFF, obtains a performance of 97.87% accuracy. It shows that the proposed work performance is improved compared to the existing methods.</p> Conclusion <p>The 3DMSFF proposed architecture proves that it is a strong and effective deep learning framework for DBT classification to combat the constraints introduced by the heterogeneity of the lesion and confounding structures through multi-scale feature learning. The finding suggests that integrating local fine-grained detail and global observation can significantly improve computer-aided breast cancer diagnosis.</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

Multiscale feature fusion for automated classification of digital breast tomosynthesis volumes using 3D deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月8日

查看全文
  • Harsha Latha P,
  • S. Ravi,
  • Manusha C,
  • Abhishek Kumar,
  • Pramod Singh Rathore

摘要

Background

Digital breast tomosynthesis (DBT) offers a better ability to visualise lesions than conventional mammography and is becoming a valuable imaging modality for detecting breast cancers. Although DBT reduces tissue overlap compared to conventional 2D mammography, automated diagnosis remains challenging due to volumetric complexity and variability in lesion appearance.

Objective

We present a novel approach, Three-dimensional Multi-scale Features Fusion (3DMSFF), for classifying DBT volumes into benign and malignant classes according to a fusion of deep learning models with a new multi-scale features fusion strategy. The new strategy combines two deep learning models to extract global context knowledge and localised fine-grained features essential for precise tumour classification.

Methodology

The proposed model consists of a two-stream 3D deep learning model to learn complementary information from 3D volumes of DBT; (i) Local Feature Extraction - Convolutional Neural Networks (LFE-CNN) is a Convolutional Neural Network (CNN) branch that extracts low-level information like textural and structural information with residual connections, (ii) Global Feature Extraction – Convolutional Neural Networks (GFE-CNN) uses separable convolutions and squeeze and excitation (SE) blocks in depth to highlight large-scale discriminative features and contextual correlations within 3D breast volume. To reduce redundancy and computational load, each DBT volume is first processed through entropy-based ranking to extract the 20 most informative slices, which are then stacked to form a refined 3D subvolume used as model input. The features learnt from the two branches are combined through a feature concatenation technique, and fully connected layers are employed to make the final classification. The model is trained with focal loss to address the class imbalance issue.

Results

The 3DMSFF architecture is evaluated on a dataset of 3D DBT volumes labelled benign or malignant. The results present a significant improvement in classification performance compared to the standard 3D CNN architecture, as testified by the increased accuracy, sensitivity, and Area Under the Receiver Operating Characteristic curve (AUC-ROC) performance. The proposed work 3DMSFF, obtains a performance of 97.87% accuracy. It shows that the proposed work performance is improved compared to the existing methods.

Conclusion

The 3DMSFF proposed architecture proves that it is a strong and effective deep learning framework for DBT classification to combat the constraints introduced by the heterogeneity of the lesion and confounding structures through multi-scale feature learning. The finding suggests that integrating local fine-grained detail and global observation can significantly improve computer-aided breast cancer diagnosis.