In this article, we propose a segmentation algorithm for skin lesions in 3D high-frequency ultrasound images. The segmentation is done on melanoma and Basal-cell carcinoma tumors, the most common skin cancer types, and could be used for diagnosis and surgical excision planning in a clinical context. Compared with previously proposed algorithms, which tend to underestimate the size of the lesion, we propose two new boundary terms which provide significant improvements of the accuracy. The first is a probabilistic boundary expansion (PBE) term designed to broaden the segmented area at the boundaries, which uses the feature asymmetry criterion. The second is a curvature dependent regularization (CDR), which aims at overcoming the tendency of standard regularization to shrink segmented areas. On a clinical dataset of 12 patients annotated by a dermatologist, the proposed algorithm has a comparable Dice index but increases the sensitivity by . The proposed algorithm improves the sensitivity for all lesions, and the obtained sensitivity is close to that of the intra-observer variability.