A deep convolutional neural network for the detection of polyps in colonoscopy images

https://doi.org/10.1016/j.bspc.2021.102654Get rights and content

Highlights

  • Provides a deep convolutional neural network-based model for the computerized detection of polyps within colonoscopy images. The proposed model comprises 16 convolutional layers with 2 fully connected layers, and a SoftMax layer.

  • We applied two different activation functions, MISH and rectified linear unit activation functions for deeper propagation of information and self-regularized smooth non-monotonicity.

  • Furthermore, we used a generalized intersection of union, thus overcoming issues such as scale invariance, rotation, and shape.

  • Detailed benchmarked results are provided, showing better performance in terms of precision, sensitivity, F1-score, F2-score, and Dice-coefficient, thus proving the efficacy of the proposed model.

Abstract

Colonic polyps detection remains an unsolved issue because of the wide variation in the appearance, texture, color, size, and appearance of the multiple polyp-like imitators during the colonoscopy process. In this paper, a deep convolutional neural network (CNN) based model for the computerized detection of polyps within colonoscopy images is proposed. The proposed deep CNN model employs a unique way of adopting different convolutional kernels having different window sizes within the same hidden layer for deeper feature extraction. A lightweight model comprising 16 convolutional layers with 2 fully connected layers (FCN), and a Softmax layer as output layer is implemented. For achieving a deeper propagation of information, self-regularized smooth non-monotonicity, and to avoid saturation during training, MISH as an activation function is used in the first 15 layers followed by the rectified linear unit activation (ReLU) function. Moreover, a generalized intersection of the union (GIoU) approach is employed, overcoming issues such as scale invariance, rotation, and shape encountering with IoU. Data augmentation techniques such as photometric and geometric distortions are employed to overcome the scarcity of the data set of the colonic polyp. Detailed experimental results are provided that are bench-marked with the MICCAI 2015 challenge and other publicly available data set reflecting better performance in terms of precision, sensitivity, F1-score, F2-score, and Dice-coefficient, thus proving the efficacy of the proposed model.

Introduction

Early diagnosing of distinct diseases within the small intestine is a time-consuming and hectic process for physicians. This has led to the introduction of technologies such as colonoscopy and wireless capsule endoscopy [1] where several images are generated during those processes. Colorectal cancer (CRC) is the second-highest cause of death by cancer worldwide with 880,792 deaths and a mortality rate of 47.60% in 2018 reported by American Cancer Society [2]. A critical step in every CRC screening program is colonoscopy, where the aim is to classify and detect pre-malignant or malignant polyps employing a camera that is inserted into the large bowel [3]. 95.00% of CRC cases begin with the appearance of a growth on the inner lining of the rectum or colon, called a polyp. Various types of polyps exist including, adenoma polyps, which can worsen into CRC. CRC is curable in 90.00% of cases considering early detection [4]. Colonoscopy has emerged as a minimally invasive and additional tool for investigating polyps by examining the gastrointestinal tract [4]. The process of colonoscopy relies on highly experienced endoscopists, where recent clinical investigations have shown that the colonoscopy process misses 22.00–28.00% of polyps. These false negatives can lead to late diagnosis of colon cancer, resulting in a survival rate as low as 10.00% [5].

DL plays an important role in many areas, including text recognition tasks, self-driving cars, image recognition, and healthcare. Computer vision and ML-based techniques have evolved over several decades and are being employed for the automatic detection of polyps [6], [7], [8]. The candidate boundaries features of polyps are obtained using low-level image processing techniques such as hand-crafted features like texture, histograms of oriented gradients, color wavelets, and local binary patterns analysis [9], [10]. Recently, advanced algorithms have been suggested to evaluate polyp appearance based on factors such as context information [11] and edge shape [12]. However, the decrease in detection performance is mainly due to the similar appearance of polyp-like and polyp structures that need to be addressed.

Convolutional neural networks (CNNs) display promising results when it comes to object detection and segmentation. In the 2015 MICCAI challenge, CNN outperformed techniques based on low image processing, i.e. hand-crafted features analysis for the detection of polyp [6]. In the last decade, the region-based CNN approaches, such as R-CNN [13], Fast R-CNN [14], and Faster R-CNN [15] presented a promising results for objects and polyps detection. Regression-based attempts are performed by employing a single-shot multi-box detector (SSD) [16] and You Only Look Once (YOLO) [17] for the detection of polyps. Despite CNN's robustness and high detection efficiency, recent investigations have shown that deep neural networks (DNNs), including CNN's, are extremely vulnerable to noise up to one single-pixel [18] and perturbations [19] can lead to miss detection. Even though computer-aided detection techniques can effectively classify and detect, polyps detection remains challenging due to its significant size, appearance, and intensity variations within the small bowel and consecutive frames. Moreover, the artifacts generated due to the existence of intestinal content (fecal and bubble particles), as well as the appearance of specular overexposed areas and highlights, may extra aggravate the circumstance. This is a serious issue, because polyps and polyp-like objects have similar appearances in consecutive frames, leading to miss-detection even when implementing powerful models such as CNN. Furthermore, the performance of DL approaches is profoundly associated with the amount of data available for training. The lack of availability of labeled polyp images for training makes the detection and segmentation of the polyp a difficult task [20].

Considering the above critical issues related to a DL-based model for the detection of the polyp, this work presents a deep CNN-based detection model of polyp in colonoscopy images. The proposed deep CNN model employs a unique way of adopting different convolutional kernels having different window sizes within the same hidden layer for deeper feature extraction. A lightweight model comprising 16 convolutional layers with 2 fully connected layers (FCN), and a Softmax layer as output layer is implemented. For achieving a deeper propagation of information, self-regularized smooth non-monotonicity, and to avoid saturation during training, MISH [21] as an activation function is used in the first 15 layers followed by the rectified linear unit activation (ReLU) function. Moreover, a generalized intersection of the union (GIoU) [22] approach is employed, overcoming issues such as scale invariance, rotation, and shape encountering with IoU. The rest of the paper is categorized as follows: Section 2 presents recent related work done for polyp detection in colonoscopy images. Section 3 details the proposed deep CNN model for polyp detection in colonoscopy images while in Section 4, the experimental results are described in detail, along with the data set specifications and data augmentation process. Finally, in Section 5 the paper is wrapped with a conclusion with the future work presentation.

Section snippets

Related work

From the past two decades, techniques based on computer vision and machine learning (ML) have been introduced for the computerized detection of polyps [19]. In initial investigations, hand-craft features, such as texture, Haar, color wavelet, the histogram of oriented gradients (HoG), and local binary pattern (LBP) were studied for the detection of polyps [23], [24], [25]. More advanced algorithms were introduced; where edge shape and context information was used in the former while valley

Proposed deep convolutional neural network (CNN) architecture

At first, the input image is divided into grid cells G×G during the training phase. The size of the grid cells for the proposed implementation is kept as 8×8 resulting in 64 cells, where each grid cell is responsible for the probability of the presence of the object within the image. Each grid cell is associated with a distinct region of the image, and these cells predict objects whose centers lay in the region. This allows the network to have a structured output description to utilize the

Experimental results and discussion

This section details the data set specifications and experimental results generated by implementing the proposed deep CNN for the detection of polyps in colonoscopy images.

Conclusion

In this paper, we presented a computerized DL-based detection model for colonic polyps. a deep convolutional neural network (CNN) based model for the computerized detection of polyps within colonoscopy images is proposed. The proposed deep CNN model employs a unique way of adopting different convolutional kernels having different window sizes within the same hidden layer for deeper feature extraction. A lightweight model comprising 16 convolutional layers with 2 fully connected layers (FCN),

Acknowledgments

This work was supported by Priority Research Centers Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology” (2018R1A6A1A03024003). This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the Grand Information Technology Research Center support program (IITP-2021-2020-0-01612) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation).

Conflict of interest:

References (44)

  • J. Bernal et al.

    Comparative validation of polyp detection methods in video colonoscopy: results from the MICCAI 2015 endoscopic vision challenge

    IEEE Trans. Med. Imaging

    (2017)
  • N. Tajbakhsh, S.R. Gurudu, J. Liang, System and methods for automatic polyp detection using convulutional neural...
  • W. Wang et al.

    An improved deep learning approach and its applications on colonic polyp images detection

    BMC Med. Imaging

    (2020)
  • N. Tajbakhsh et al.

    Automated polyp detection in colonoscopy videos using shape and context information

    IEEE Trans. Med. Imaging

    (2015)
  • R. Girshick et al.

    Rich feature hierarchies for accurate object detection and semantic segmentation

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    (2014)
  • R. Girshick

    Fast r-cnn

    Proceedings of the IEEE International Conference on Computer Vision

    (2015)
  • S. Ren et al.

    Faster r-cnn: towards real-time object detection with region proposal networks

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2016)
  • W. Liu et al.

    Ssd: single shot multibox detector

  • J. Redmon et al.

    Yolov3: An Incremental Improvement

    (2018)
  • J. Su et al.

    One pixel attack for fooling deep neural networks

    IEEE Trans. Evol. Comput.

    (2019)
  • H.A. Qadir et al.

    Improving automatic polyp detection using cnn by exploiting temporal dependency in colonoscopy video

    IEEE J. Biomed. Health Informatics

    (2019)
  • W.-L. Chao et al.

    Application of artificial intelligence in the detection and differentiation of colon polyps: a technical review for physicians

    Diagnostics

    (2019)
  • Cited by (0)

    View full text