Feasibility Study of Lesion Detection Using Deformable Part Models in Breast Ultrasound Images

Detection of lesions in ultrasound imaging typically requires human analysis due to their complexity. Hence, computerized lesion detection methods could be used to help radiologists in this process due to the fact that an early detection reduces the death rate caused by breast cancer. In this paper we propose a first experiment of a feasibility study for adapting a generic object detection technique, Deformable Part Models (DPM), to detect lesions in breast US images without any kind of human supervision. This technique has been evaluated in different topics obtaining prominent results. Hence, we propose a first assessment of this methodology applied to lesion detection in US images. We used a data-set composed by 50 images, all from different patients (18 malignant lesions, 32 benign lesions and 50 healthy tissue regions). In terms of quantitative results for lesion detection, our proposal obtains a sensitivity of 82% with 0.51 false-positive detections per image and an A(z) value of 0.96, which proves the feasibility of the proposal.