Deep Learning for breast ultrasound analysis:

a CNN-based tumor segmentation and classification for improved diagnosis

Abstract

Breast ultrasound imaging is an essential tool in early breast cancer detection, yet its interpretation remains a challenging task due to image variability and noise. This study explores deep learning-based approaches for tumor segmentation and classification in breast ultrasound images, aiming to improve diagnostic accuracy and assist medical professionals in decision-making. An encoder-decoder architecture utilizing two pre-trained convolutional neural networks, DeepLabV3+ and U-Net, is proposed for the segmentation task. The segmentation performance was evaluated against a semi-automatic Local Graph Cut method using the Dice similarity coefficient. DeepLabV3+ achieved superior results compared to U-Net and Local Graph Cut.
Further, a deep learning framework incorporating MobileNetV2, VGG16, and EfficientNetB7 is employed for
classification. The proposed approach is novel in its ability to extract and analyze features from both the lesion and the surrounding tissue, leveraging morphological operations (erosion and dilation) to improve the model’s interpretability. Transfer learning allows for the optimization of classification performance. The system was trained and validated using the BUS-BRA and BUSI datasets. High accuracy and AUC scores were achieved for the classification of both benign and malignant lesions. These results confirm the effectiveness of CNNs in segmentation and classification tasks, highlighting the potential of deep learning for automated breast cancer diagnosis. The proposed methodology paves the way for more robust, interpretable, and clinically relevant AI driven diagnostic tools in breast imaging.