Volume XLII-2/W12
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2/W12, 231-235, 2019
https://doi.org/10.5194/isprs-archives-XLII-2-W12-231-2019
© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2/W12, 231-235, 2019
https://doi.org/10.5194/isprs-archives-XLII-2-W12-231-2019
© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.

  09 May 2019

09 May 2019

DETECTING PROTRUSION LESION IN DIGESTIVE TRACT USING A SINGLE-STAGE DETECTION METHOD

L. Wang1, S. Wang1, S. Huang1, and C. Liu2 L. Wang et al.
  • 1Department of Computer Science, School of Information Science and Engineering, Xiamen University, Xiamen 361005, China
  • 2Department of Medical Imaging, The Chenggong Hospital Affiliated to Xiamen University, Xiamen 361005, China

Keywords: Protrusion Lesion, Deep Learning, Single-stage Method, Multi-scale Feature Layers

Abstract. The classification networks have already existed for a long time and achieve great success. However, in biomedical image processing, classifying normal and abnormal ones only is not enough clinically, the desired output should include localization, i.e., where the lesion is. In this paper, we present a method for detecting protrusion lesion in digestive tract. We use a deep learning-based model to build a computer-aided diagnosis system to help doctors examine the intestinal diseases. Learn from existing detection method, one-stage and two-stage detection algorithm, a new network suitable for protrusion lesion detection is proposed. We inherit the method of anchor generation in SSD, a fast single-stage object detector outperform R-CNN series in terms of speed. Multi-scale feature layers are assigned to generate different sizes of default anchor boxes. Different from the previous work, our method doesnt require additional preprocessing because the network can learn features autonomously. For the 256*256 input, our method achieves 73% AP, perform a novel way to detect protrusion lesions.