Laser & Optoelectronics Progress, Volume. 60, Issue 6, 0617002(2023)

Improved YOLOv4 Model-Based Spinal Magnetic Resonance Imaging Image Detection

Ning Dai1、*, Yuhai Gu1, Zhicheng Zhang2, Yang Zhang2, and Zhan Xu1
Author Affiliations
  • 1Key Laboratory of Modern Measurement and Control Ministry of Education, Beijing Information Science and Technology University, Beijing 100192, China
  • 2Department of Orthopedics, PLA General Hospital, Beijing 100700, China
  • show less
    Figures & Tables(18)
    YOLOv4 network structure
    Mish activation function
    MRI image of spine
    Histogram of gray value before and after transformation. (a) Original drawing; (b) histogram equalization; (c) CLAHE
    Images before and after equalization. (a) Original drawing; (b) histogram equalization; (c) CLAHE
    Average RIOU of different anchor frames
    Depthwise separable convolution
    Optimized YOLOv4 network structure
    Loss curves. (a) YOLOv4; (b) YOLOv4-disc
    YOLOv4-disc P-R curve. (a) Healthy; (b) protrusion; (c) swollen
    AP values of different networks
    Comparison of missed inspection. (a) YOLOv4; (b) YOLOv4-disc
    Comparison of repeated detection. (a) YOLOv4; (b) YOLOv4-disc
    Comparison of nonoptimal shooting positions. (a) YOLOv4; (b) YOLOv4-disc
    • Table 1. Feature label

      View table

      Table 1. Feature label

      Feature labelNumber
      Ptrotrusion6410
      Swollen3846
      Healthy2564
    • Table 2. Verification results of different network models

      View table

      Table 2. Verification results of different network models

      NetworkHealthyProtrusionSwollenSize /MBFPS
      Precision /%Recall /%Precision /%Recall /%Precision /%Recall /%
      YOLOv4-disc95.1893.5497.3496.7182.4681.8228134.7
      YOLOv492.6191.1692.3591.8482.4281.6135428.5
      YOLOv383.4278.6582.7782.1374.3670.2523730.9
      Faster-RCNN81.3580.2280.6779.4570.7568.9214627.6
    • Table 3. F1 values for different networks

      View table

      Table 3. F1 values for different networks

      NetworkHealthy-F1Protrusion-F1Swollen-F1
      YOLOv4-disc0.940.970.82
      YOLOv40.910.920.81
      YOLOv30.810.820.72
      Faster-RCNN0.800.800.69
    • Table 4. AP values of different networks

      View table

      Table 4. AP values of different networks

      NetworkHealthyProtrusionSwollenmAP
      YOLOv4-disc94.4896.2181.7390.80
      YOLOv491.1792.2881.4287.29
      YOLOv382.0183.3472.7779.37
      Faster-RCNN79.5879.7268.4875.92
    Tools

    Get Citation

    Copy Citation Text

    Ning Dai, Yuhai Gu, Zhicheng Zhang, Yang Zhang, Zhan Xu. Improved YOLOv4 Model-Based Spinal Magnetic Resonance Imaging Image Detection[J]. Laser & Optoelectronics Progress, 2023, 60(6): 0617002

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Medical Optics and Biotechnology

    Received: Nov. 25, 2021

    Accepted: Jan. 11, 2022

    Published Online: Mar. 16, 2023

    The Author Email: Ning Dai (18501301242@163.com)

    DOI:10.3788/LOP213059

    Topics