Breast Tumor Classification using Short-ResNet with Pixel-based Tumor Probability Map in Ultrasound Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Biomedical Engineering

Background:

• Breast cancer is a leading cause of death for women globally.
• Early detection and accurate diagnosis of breast cancer are critical for improving patient outcomes.
• Breast ultrasound is a key imaging modality, but accurate segmentation and classification of tumors remain challenging.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for simultaneous segmentation and classification of breast tumors in ultrasound images.
• To improve the accuracy and efficiency of breast cancer diagnosis using automated image analysis.
• To address the challenges in precise tumor delineation and benign/malignant differentiation in breast ultrasound.

Main Methods:

• A hybrid deep learning architecture, short-ResNet combined with DC-UNet, was proposed for segmentation and classification.
• The model was trained and validated on breast ultrasound datasets.
• Performance was evaluated using metrics such as Dice coefficient for segmentation and accuracy for classification.

Main Results:

• The proposed model achieved an 83% Dice coefficient for breast tumor segmentation.
• The model demonstrated a 90% accuracy in classifying breast tumors as benign or malignant.
• Comparative experiments on different datasets confirmed the model's generalizability and superior performance.

Conclusions:

• The short-ResNet with DC-UNet model effectively addresses the challenges of breast tumor segmentation and classification in ultrasound images.
• This deep learning approach shows significant potential for enhancing early breast cancer detection and diagnosis.
• The combined segmentation and classification strategy improves diagnostic accuracy and aids in clinical decision-making.