Binary Seeds Auto Generation Model for Knee Cartilage Segmentation

Segmentation is an instrumental task in medical image analysis. In addition to existing manual, semi-automatic and automatic segmentation models, deep learning has been the niftiest machine learning technique in current research interests. However, none of the models or technique can escape from the overdependence on training data and user intervention. As a result, the use of computer-aided and learning algorithms have reported lackluster robustness in the presence of high anatomical disparity. In recognition of this, we have proposed a binary seeds auto-generation model to reduce the reliance on manually crafted priori information in deep learning. Then, we computed the reproducibility of the proposed model against manual segmentation using normal and osteoarthritic knee magnetic resonance image. In normal knee image, mean agreements of the proposed model and manual segmentation were 0.94 ± 0.022 and 0.83 ± 0.028 respectively. In osteoarthritic knee image, mean agreements of the proposed model and manual segmentation were 0.92 ± 0.051 and 0.79 ± 0.073 respectively. Pair t test showed that our method has better accuracy than manual segmentation in both cases (normal: P=1.03×10
^-9
; osteoarthritic: P=4.94×10
^-6
). Therefore, we can conclude the model is robust to be implemented as part of deep learning based segmentation framework.