다중 스케일 용종 영상 특징 맵간 의미론적 간극을 줄이는 양방향 융합 네트워크 = Bi-directional Fusion Network to Reduce Semantic Gaps between Multi-scale Polyp Feature Maps

국문 초록 (Abstract)

대장내시경은 대장 내 점막에 있는 용종을 탐지하고 제거하는데 있어 현존하는 의학 기술 중 가장 효과적인 시술이다. 대장 내시경 검사 중 용종 검출 성능을 향상시키기 위한 용종 영상 분할 심층신경망 모델들이 제안되어왔다. 하지만, 용종은 다양한 크기와 형태를 지니고 주변 점막과 뚜렷하지 않은 경계를 가지고 있어 자동화된 용종 분할 시 일관적인 성능을 보장하기 어렵다. 용종 영상을 효과적으로 분할하기 위해서는 다양한 크기의 수용영역을 가진 특징 맵을 활용하는 것이 필요한데, 서로 다른 크기로 인해 의미론적 간극이 존재하는 특징 맵들을 융합하는 과정에서 정보 손실이 생길 수 있다. 이러한 문제를 해결하기 위해 본 논문에서는 다중 스케일에 강한 특징 맵 융합 및 정제 네트워크인 MPISNet (Multi-scale Polyp Image Segemtnation Network)을 제안한다. MPISNet에서는 새로운 융합 모듈인 SCBi-FPN (Semantically Calibrated Bi-directional Feature Pyramid Network)을 활용하여 특징 맵 간 의미론적 간극을 줄이고, 기존의 합성곱 신경망 주의 모듈인 CBAM (Convolutional Block Attention Module)을 활용하여 정제하였다. 결과적으로, Kvasir-SEG 데이터셋에서 기존 PraNet 모델에 비해 F1 스코어, 재현율, mIoU에서 각각 1.72%, 1.15%, 2.13%의 성능 향상을 보였으며 180fps의 합리적인 추론 속도를 보였다.

다국어 초록 (Multilingual Abstract)

Colonoscopy is the gold standard method for detecting colorectal polyps that are likely to progress into colorectal cancer. Many deep learning models have been suggested to aid and assist doctors in detecting and removing potentially harmful polyps during colonoscopy. However, segmenting polyp images can be a challenging task due to polyps’ variation in size, shape, and color and their similarity to the surrounding mucosal area. In order to effectively segment polyps, it is essential to fuse feature maps of various sizes while considering the semantic gaps in between. To tackle this problem, we propose MPISNet (Multi-scale Polyp Image Segemtnation Network) which is designed to detect polyps that vary in size by filling in the semantic gaps between features produced with different receptive fields. Specifically, a novel SCBi-FPN module was used to effectively fuse features of different sizes and CBAM module to re-calibrate feature maps using both channel and spatial attentions. Our network outperforms PraNet, a previous network for polyp segmentation, in Kvasir-SEG dataset by 1.72%, 1.15%, 2.13%, in F1-score, Recall, and mIoU, while showing a reasonable mean inference rate of 180fps.

참고문헌 (Reference)

• 1 Zongwei Zhou, "UNet++ : A Nested U-Net Architecture for Medical Image Segmentation" 2018
• 2 Olaf Ronneberger, "U-Net : Convolutional Networks for Biomedical Image Segmentation" 2015
• 3 Silva J. Histace, "Toward embedded detection of polyps in wce image for early diagnosis of colorectal cancer" 2014
• 4 Jie Hu, "Squeeze-and-Excitation Networks" 2018
• 5 Vijay Badrinarayanan, "SegNet : A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation" IEEE 2017
• 6 Jinke Wang, "SAR-U-Net : Squeeze-and-excitation block and atrous spatial pyramid pooling based residual U-Net for automatic liver segmentation in Computed Tomography" 2021
• 7 Debesh Jha, "ResUNet++: An Advanced Architecture for Medical Image Segmentation" IEEE 2019
• 8 Foivos I. Diakogiannis, "ResUNet : a deep learning framework for semantic segmentation of remotely sensed data" 2019
• 9 Shang-Hua Gao, "Res2Net : A New Multi-scale Backbone Architecture" 2019
• 10 Huisi Wu, "Precise Yet Efficient Semantic Calibration and Refinement in ConvNets for Real-time Polyp Segmentation from Colonoscopy Videos" 2021
• 11 Deng-Ping Fan, "PraNet : Parallel Reverse Attention Network for Polyp Segmentation" 2020
• 12 Sebastian Gross, "Polyp Segmentation in NBI Colonoscopy" 2009
• 13 Lequan Yu, "Integrating Online and Offline Three-Dimensional Deep Learning for Automated Polyp Detection in Colonoscopy Videos" IEEE 2016
• 14 Jonathan Long, "Fully Convolutional Networks for Semantic Segmentation" 2015
• 15 Tsung-Yi Lin, "Feature Pyramid Networks for Object Detection" 2017
• 16 Leufkens, "Factors influencing the miss rate of polyps in a backto-back colonoscopy study" 44 (44): 2012
• 17 Liang-Chieh Chen, "Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation" 2018
• 18 Mingxing Tan, "EfficientDet : Scalable and Efficient Object Detection" 2020
• 19 Frank T. Kolligs, "Diagnostics and Epidemiology of Colorectal Cancer" 2016
• 20 Sánchez-Peralta, "Deep learning to find colorectal polyps in colonoscopy : A systemic literature review" 2020
• 21 Hiroyuki Yoshida, "Computer-aided Diagnosis Scheme for Detection of Polyps at CT Colonography" 2002
• 22 Zaiwang Gu, "CE-Net : Context Encoder Network for 2D Medical Image Segmentation" IEEE 2019
• 23 Sanghyun Woo, "CBAM : Convolutional Block Attention Module" 2018
• 24 Nima Tajbakhsh, "Automated Polyp Detection in Colonoscopy Videos Using Shape and Context Information" IEEE 2015