Deep Learning-Based Breast Tumor Classification Using Shear-Wave Sonoelastography Image Features and Clinical...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Shear-wave elastography (SWE) aids in early and efficient breast mass classification.
• Convolutional Neural Networks (CNNs) show promise in analyzing medical images.
• Integrating clinical data with imaging features can potentially improve diagnostic accuracy.

Purpose of the Study:

• To evaluate the role of clinical variables and SWE image features in breast mass classification.
• To assess the performance of various CNN models (ResNet101, VGG16, Exception, InceptionV3, DenseNet169) in this classification task.
• To determine the impact of region of interest (ROI) selection on classification performance.

Main Methods:

• Prospective collection of 834 SWE images; 534 used for training CNNs.
• CNN models trained using image features alone and combined with clinical variables.
• Evaluation of classification performance with and without manual ROI selection on B-mode images.

Main Results:

• DenseNet169 and ResNet152 achieved the highest performance when using both clinical variables and SWE image features.
• DenseNet169 achieved 94.01% accuracy and AUCs of 0.86 (Test) and 0.97 (Validation).
• Combining clinical variables and image features significantly improved AUCs for DenseNet169, VGG16, and InceptionV3 with ROI selection (P ≤ 0.05).

Conclusions:

• Clinical variables significantly enhance the performance of most CNNs for breast mass classification on SWE images with ROI selection.
• The integration of clinical data alongside SWE imaging and CNN analysis offers a more robust approach to breast lesion characterization.
• CNN models, particularly DenseNet169 and ResNet152, demonstrate high efficacy in classifying breast masses when provided with comprehensive data.