Related Experiment Video
Updated: Nov 27, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.1K
Breast Tumor Classification in Ultrasound Images Using Combined Deep and Handcrafted Features
Mohammad I Daoud1, Samir Abdel-Rahman1, Tariq M Bdair2
1Department of Computer Engineering, German Jordanian University, Amman 11180, Jordan.
Sensors (Basel, Switzerland)
|December 3, 2020
Summary
Combining deep learning features from VGG19 models with handcrafted features significantly improves breast ultrasound tumor classification. This hybrid approach enhances accuracy and reduces misclassification of benign tumors.
Area of Science:
- Medical Imaging
- Computer Science
- Artificial Intelligence
Background:
- Accurate breast ultrasound image classification is crucial for early tumor detection.
- Deep learning models show promise but can benefit from integration with traditional methods.
Purpose of the Study:
- To enhance breast ultrasound image classification by combining deep and handcrafted features.
- To identify optimal feature combinations for improved tumor classification accuracy.
Main Methods:
- Extracted deep features from VGG19 model at multiple levels.
- Applied feature selection algorithms to identify best deep and combined feature sets.
- Combined deep features with handcrafted texture and morphological features.
- Validated performance using cross-validation on 380 breast ultrasound images and tested generalization on 163 additional images.
Main Results:
- The CONV features (deep features from all VGG19 convolution blocks) achieved 94.2% accuracy.
- Combining CONV features with handcrafted morphological features improved performance to 96.1% accuracy.
- The hybrid approach outperformed handcrafted features and a fine-tuned VGG19 model.
- Generalization was confirmed on an independent dataset.
Conclusions:
- A combination of deep VGG19 features and handcrafted morphological features offers superior breast ultrasound image classification.
- This method improves malignant tumor detection and reduces benign tumor misclassification.
- The findings support the clinical utility of hybrid feature approaches in medical imaging analysis.
