Accurate detection of large-scale, elliptical-shape fibers, including their parameters of center, orientation and major/minor axes, on the 2D cross-sectioned image slices is very important for characterizing the underlying cylinder 3D structures in microscopic material images. Detecting fibers in a degraded image poses a challenge to both current fiber detection and ellipse detection methods. This paper proposes a new semi-supervised deep learning method for large-scale elliptical fiber detection with synthetic data, which frees people from heavy data annotations and is robust to various kinds of image degradations. A domain adaptation strategy is utilized to reduce the domain distribution discrepancy between the synthetic data and the real data, and a new Region of Interest (RoI)-ellipse learning and a novel RoI ranking with the symmetry constraint are embedded in the proposed method. Experiments on real microscopic material images demonstrate the effectiveness of the proposed approach in large-scale fiber detection.