Hybrid deep learning models for automatic segmentation and classification of breast lesions in ultrasound images

Abstract Background and purpose Breast cancer requires early detection for effective treatment. Although ultrasound offers high sensitivity and is less invasive, it still faces challenges such as speckle noise and complex tissue textures. This study aims to develop hybrid deep learning models to improve automatic breast tumor segmentation and classification, and to assess whether combining segmentation with classification models enhances performance compared to using these models individually. Methods This study employed the publicly available dataset, containing 780 images from 600 patients, classified into normal, benign, and malignant lesions. In the initial step, image preprocessing involved resizing images, normalizing pixel intensities, and performing data augmentation. Pre-trained deep learning models—VGG-16, DenseNet-121, DenseNet-169, and ResNet-50—are fine-tuned for classification tasks. Models are evaluated based on accuracy, precision, recall, F1-score, and area under the curve (AUC). The U-Net model is used for generating segmentation masks to highlight regions of interest, with performance assessed through IoU and Dice coefficient. The study compares classification performance using original versus segmented images. Results On original breast ultrasound images, DenseNet-169 achieved 87% accuracy with an AUC of 0.99, while DenseNet-121 reached an AUC of 0.85. With segmented images, both DenseNet-121 and DenseNet-169 achieved high AUC values (≈0.99–1.00), with DenseNet-121 also reaching 98% accuracy. VGG-16 obtained an AUC of 0.99 and 97% accuracy, and ResNet-50 achieved an AUC of 0.93 with 84% accuracy. Conclusion Applying U-Net segmentation before classification improved model performance on the BUSI dataset, particularly for DenseNet-121 and DenseNet-169, by helping the classifiers focus on lesion-specific regions. However, as this work represents an initial technical evaluation using a single-center dataset, the findings should be considered preliminary. External validation on multicenter cohorts and prospective studies will be essential before any consideration of real-world deployment.