Semiautomatic contour detection of breast lesions in ultrasonic images with morphological operators and average radial derivative function

This work presents a computerized lesion segmentation technique on breast ultrasound images. There were applied known techniques such as morphological filtering, Watershed transformation and average radial derivative function. To evaluate the performance of the proposed method, two protocols were established. For the first, the resulting segmentation contours were compared with those of 24 gold standard simulated ultrasound-like images, and, for second, with 36 breast US images manually delineated by two senior radiologists. Further, two evaluation parameters were used: the percentage of coincidence (CP) and the proportional distance (PD). The former indicates the similarity between contours, while the latter express the dissimilarity. The accuracy of the proposed method was evaluated by considering images with CP>80% and PD<10% as adequately delineated. It was higher than 80% for real images and higher than 88% for simulated images.