DRR-Net: A Dense-Connected Residual Recurrent Convolutional Network for Surgical Instrument Segmentation From Endoscopic Images
- 25 July 2022
- journal article
- research article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Robotics and Bionics
- Vol. 4 (3) , 696-707
Abstract
The precise segmentation of surgical instruments is the key link for the stable and reasonable operation of surgical robots. However, accurate surgical instrument segmentation is still a challenging task due to the complex surgical environment in endoscopic images, low contrast between surgical instruments and tissues, and the diversity of surgical instruments and their morphological variability. In recent years, deep learning has been widely applied into medical image segmentation and achieved a certain achievements, especially U-Net and its variants. However, existing surgical instrument segmentation networks still suffer from some shortcomings, such as insufficient processing of local feature maps, lack of temporal modeling information, etc. To address the above issues, in order to effectively improve the segmentation accuracy of surgical instruments, based on an encoder-decoder network structure, a novel dense residual recurrent convolutional network, called DRR-Net, is proposed in this paper for automatic and accurate surgical instrument segmentation from endoscopic images. Faced with lack of temporal modeling information, inspired by the recurrent neural networks (RNNs), an attention dense-connected recurrent convolutional block (ADRCB) is proposed to optimize the backbone network to obtain temporal information and learn the correspond pixel relationship between frames. To address the insufficient processing issue of local feature maps, to replace the simple skip connections, a residual path is proposed to enhance the context feature representation. Meanwhile, it could also reduce the effect of semantic gap issue. Further, to improve the segmentation accuracy on segmented objects with different sizes, a context fusion block (CFB) is proposed to embed into the bottleneck layer to extract multi-scale attention context features. Multiple public data sets on surgical instrument segmentation are adopted for model evaluation and comparison, including kvasir-instrument set and UW-Sinus-Surgery-C/L set. Experimental results demonstrate that proposed DRR-Net network could achieve an excellent performance compared with other advanced segmentation models.Keywords
Funding Information
- National Key Research and Development Project of China (2020YFB1313701)
- National Natural Science Foundation of China (62003309)
- Outstanding Foreign Scientist Support Project in Henan Province of China (GZS2019008)
This publication has 32 references indexed in Scilit:
- HL-FCN: Hybrid loss guided FCN for colorectal cancer segmentation2018
- A geometric flow approach for region-based image segmentation-theoretical analysisActa Mathematicae Applicatae Sinica, English Series, 2018
- Convolutional Neural Networks2017
- Densely Connected Convolutional Networks2017
- DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFsIEEE Transactions on Pattern Analysis and Machine Intelligence, 2017
- SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image SegmentationIEEE Transactions on Pattern Analysis and Machine Intelligence, 2017
- Vision-based and marker-less surgical tool detection and tracking: a review of the literatureMedical Image Analysis, 2017
- Deep Residual Learning for Image Recognition2016
- A Miniature Soft Robotic Manipulator Based on Novel Fabrication MethodsIEEE Robotics and Automation Letters, 2016
- Fully convolutional networks for semantic segmentation2015