Use of simplex search in active shape models for improved boundary segmentation

Active shape models (ASMs) are an effective boundary segmentation technique currently applied to a variety of image analysis problems, they are fast and accurate if the initial model pose and shape are close enough to the object boundary. An ASM consists essentially of a statistical shape model (point distribution model (PDM)), and a local search method along normal pixel profiles, which are located at each point of the PDM. In this paper a new ASM fitting algorithm which incorporates the optimization of an objective function together with the local search currently used in ASMs is reported. The objective function was constructed as the mean Mahalanobis distance of all the pixel profiles of each point of the PDM, to the corresponding mean profile of a training set, and is optimized using simplex search, which provides a fast numerical optimization. Our ASM-simplex algorithm increases significantly the range of initial model poses (scale (s0), rotation (θ0), and translations (Tx0, Ty0)) which results in more accurate boundary segmentation, without the need for any additional model training. Evaluation was performed using the shape models of the prostate and the left hand. The mean maximum error of prostate segmentation (for a range of initial rotations of [θ0 − 72, θ0 + 72] degrees) in ultrasound images, decreased 16% for the ASM-simplex algorithm with respect to the original ASM. The following reductions were obtained on photographic images of the left hand for each initial pose parameter: 39% for [Tx0 − 48, Tx0 + 48] pixels, 38% for [Ty0 − 48, Ty0 + 48] pixels, 9% for [s0 − 0.25, s0 + 0.25] and 42% for [θ0 − 30, θ0 + 30] degrees.