Breast tumor segmentation in ultrasound images: comparing U-net and U-net + +

Abstract

Purpose To compare the performance of U-net and U-net + + neural networks for the segmentation of breast tumors in ultrasound (US) images, focusing on a robust and easily reproducible methodology. Methods The Breast Ultrasound Images Dataset (BUSI), comprising 780 images, was used with 60% separated for training, 20% for validation, and 20% for testing. Data augmentation (DA) techniques were applied to improve model generalization, and an Intersection over Union (IoU)–based loss function was utilized. Both models were trained for 1000 epochs. The Dice coefficient was the primary metric for evaluating segmentation performance, while accuracy, precision, recall, and F1-score were used for assessing the model’s ability to detect tumors. Results U-net + + demonstrated superior performance, achieving a maximum average Dice score of 75.71% on validation data, compared to 75.17% for U-net. On test data, U-net + + obtained a median Dice score of 88.60% with an IQR of 30.53%. In the classification task, U-net + + achieved an accuracy of 90% and a F1-score of 94% for tumor detection. Conclusions U-net + + outperforms the base U-net in segmentation task on breast US images. While the results are promising, further enhancements could be achieved with larger datasets or integration with more complex techniques. It is encouraged a more detailed hyperparameter tuning, as it could significantly improve the results.