U-NET++ BASED BREAST CANCER DETECTION

Abstract

This research investigates the application of the U-Net++ deep learning architecture for the automated segmentation of breast tumors in ultrasound images. Accurate tumor segmentation is crucial for the effective diagnosis and treatment planning of breast cancer. The methodology encompasses a detailed preprocessing pipeline, U-Net++ model implementation, and rigorous evaluation. Initially, ultrasound images and corresponding masks were resized to 196x196 pixels to standardize input dimensions. Class distribution analysis revealed an imbalance. The U-Net++ architecture, known for its nested skip connections, was employed for semantic segmentation. The encoder extracts features through convolutional blocks with ELU activation, SpatialDropout2D, and batch normalization. The decoder reconstructs the segmentation using transposed convolutions. Model training utilized the AdamW optimizer, Dice loss function, and evaluation metrics, including IoU and accuracy. Callbacks, including ModelCheckpoint and EarlyStopping, were implemented to optimize training and prevent overfitting. The model achieved the following results: Training Dataset (IoU: 0.5383, Accuracy: 0.8880, Loss: 0.5749) and Test Dataset (IoU: 0.5667, Accuracy: 0.9124, Loss: 0.5963). While the model demonstrates promise, the moderate IoU scores suggest that further refinement is necessary for enhanced segmentation precision and clinical applicability.