An Auto-Contouring Method for Kidney Using a Novel Semi-Supervised Leaning Extreme Learning Machine Method

Abstract

Background: Adaptive radiation therapy planning requires contour segmentation of dangerous organs in medical images. However, manual contour rendering is the most time-consuming and laborious work in radiotherap planning. In order to solve this problem, we propose a novel semi-supervised leaning extreme learning machine (SSL-ELM) method to realize abdominal Magnetic Resonance Imaging (MRI) guided Adaptive Radiation Therapy (MR-ART) automatic contour rendering. Method/Material: Our algorithm is based on the assumption that data within the same class are close to each other. We use this heuristic method to improve the ELM algorithm. The experimental results show that our proposed method outperforms existing classification algorithms. We used a data set of eight patients with unresectable abdominal malignant tumors recruited by professionals approved by the Cleveland Medical Center Institution Review Board. Each group included MRI and airborne T1MRI with randomly selected treatment phases. Each MRI consisted of 16 slices with a resolution of 370×370 pixels. Manual and automatic contours of the kidney were compared using Dice similarity index (DSI). Results: The proposed SSL-ELM algorithm has better performance than most classification algorithms, and the experimental results also show that the DSI values are above 0.87, with some samples reaching 0.99.