Local mean based adaptive thresholding to classify the cartilage and background superpixels

Semi-automatic segmentation is common in medical image processing because anatomical geometries demonstrated by human anatomical parts often requires manual supervision to provide desirable results. However, semi-automatic segmentation has been infamous for requiring excessive human intervention and time consuming. In order to reduce aforementioned problems, seed labels have been generated automatically using superpixels in our previous works. A fixed threshold method has been implemented to classify cartilage and background superpixels but this method is reported to lack the adaptiveness to changing image properties in 3D magnetic resonance image of knee. As a result, the coverage of background seeds are not sufficient to cover whole background area in some cases. In this work, we proposed a local mean based adaptive threshold method as a better alternative to the fixed threshold method. We calculated local mean for each block in an integral image and then use it to differentiate background superpixels from cartilage superpixels. The method is robust to illumination changes and simple to use. We tested the adaptive threshold on 35 knee images of different anatomical geometries and proved the proposed method could provide more comprehensive background seed labels distribution compared to fixed threshold method.