U-Net transfer learning backbones for lesions segmentation in breast ultrasound images

<p>Breast ultrasound images are highly valuable for the early detection of breast cancer. However, drawback these is low-quality resolution and presence speckle noise, which affects their interpretability makes them radiologists’ expertise-dependent. As medical images, datasets scarce imbalanced, annotating tedious time-consuming. Transfer learning, as a deep learning technique, can be used to overcome dataset deficiency in available images. This paper presents implementation transfer U-Net backbones automatic segmentation lesions implements threshold selection mechanism deliver optimal generalized results tumors. The work uses public (BUSI) ten state-of-theart candidate models backbones. We have trained with five-fold cross-validation technique on 630 benign malignant cases. Five out showed good results, best backbone was found DenseNet121. It achieved an average Dice coefficient 0.7370 sensitivity 0.7255. model’s robustness also evaluated against normal cases, model accurately detected 72 113 higher than four models.</p>