Colorectal cancer is one of fatal cancer worldwide. Colonoscopy is the standard treatment for examination, localization, and removal of colorectal polyps. However, it has been shown that the miss-rate of colorectal polyps during colonoscopy is between 6 to 27% [1]. The use of an automated, accurate, and real-time polyp segmentation during colonoscopy examinations can help the clinicians to eliminate missing lesions and prevent further progression of colorectal cancer. The “Medico automatic polyp segmentation challenge” provides an opportunity to study polyp segmentation and build a fast segmentation model. The challenge organizers provide a Kvasir-SEG dataset to train the model. Then it is tested on a separate unseen dataset to validate the eficiency and speed of the segmentation model. The experiments demonstrate that the model trained on the Kvasir-SEG dataset [5] and tested on an unseen dataset achieves a dice coeficient of 0.7801, mIoU of 0.6847, recall of 0.8077, and precision of 0.8126, demonstrating the generalization ability of our model. The model has achieved 80.60 FPS on the unseen dataset with an image resolution of 512 × 512.
Colorectal cancer is one of the dangerous types of cancer, adding to significant deaths worldwide. Polyps are an early indicator of this type of cancer, and clinicians often detect it through colonoscopy. These polyps come in various shapes and sizes and are sometimes missed by clinicians as some polyps are hard to diferentiate from the surrounding tissue. Sometimes these polyps are covered with stool, mucosa, and other surrounding structures and pose challenges for clinicians. This is why it is essential to build a ComputerAided Diagnosis (CADx) system for detecting polyps.
The automatic polyp segmentation can play a crucial role in identifying and localizing the afected regions from the images or video frames. Semantic segmentation helps you analyze each pixel and classify them into a well-defined polyp or non-polyp class instance. With the increase in the amount of publicly available datasets, dominant methodology such as convolutional neural networks and improved hardware enables researchers to solve the challenging task of automated diagnosis of colorectal cancer in real-time.
The “Medico Automatic Polyp Segmentation Challenge” [
The proposed architecture is a full convolution network following
an encoder-decoder approach. It combines the strength of residual
learning [
The residual block consists of two 3 × 3 convolutions, where the ifrst convolution is followed by a Batch Normalization (BN) and a Rectified Linear Unit (ReLU) activation function. Next comes the second convolution layer with a BN and a squeeze and excitation layer, which is then added with the residual block’s input. After that, a ReLU activation function is used, which acts as the output of the residual block. The residual block helps us in building deep neural networks by solving the vanishing gradient and exploding gradient problem.
The convolution operation takes an image to apply a number of iflters generating a new set of output features maps. The squeeze and excitation layer provides channel-wise attention to these feature maps, strengthening the important feature and suppressing irrelevant features. The squeeze and excitation is two steps approach. At first, the incoming feature maps are first squeezed by using a global average pooling operation. This helps to get a global feature vector of size , where represents the number of feature channels from the incoming feature maps. In the second step(excitation), this global feature vector goes through a two-layer feed-forward neural network. Here the number of features is first reduced and then expanded to the original size . Finally, a sigmoid activation function is used, which scales the feature vector value between 0 and 1. This scaled feature vector is used to multiply the incoming feature maps.
The proposed network takes the polyp image of 512 × 512 size as the input given to the first encoder block. Each encoder block starts with two residual blocks, where each residual block consists of two 3 × 3 convolution and a shortcut connecting the output of both the convolution layer, also called the identity mapping. The output of the second residual block acts as the skip connection of the corresponding decoder block. These skip connections provide early information to the decoder block, which improves feature reconstruction and better model performance. It is followed by a 2 × 2 max-pooling operation, which downscales the incoming feature maps. This process is repeated 4 times, thereby reducing the feature maps size by 8 times (64 × 64) the original input size.
ReLU 3 x 2 1 5 x 2 1 5 32 64
C o n c a t e n a t e 16 144 1 x 2 1 5 x 2 1 5 128 256 128 64 32
The output of the last encoder block acts as the input of the first decoder block. In each decoder block, first, the incoming feature map is upscaled by using a 4 × 4 transpose convolution. After that, the upscaled feature map is concatenated with the appropriate same resolution feature map from the encoder via skip-connection. The skip connection provides information that is sometimes lost due to the depth of the network. It helps the decoder in the better reconstruction of the semantic feature maps. Next, follows the two residual blocks assisting the network in to learn necessary features via back-propagation. This process is repeated four times, thereby getting the output feature map of size 512 × 512. Now, the last three decoder blocks’ output is taken and upscaled to the resolution of 512 × 512 using a 4 × 4 transpose convolution. Next, we concatenate these three upscaled feature maps and the skip-connection feature maps from the first encoder block. These concatenated feature maps are then passed through a 1 × 1 convolution and sigmoid activation function. The output of the sigmoid activation results in a binary mask. 3
Table 1 shows the overall results on the validation dataset of
KvasirSEG and unseen test dataset provided by the challenge organizers.
For the evaluation of the results, the Mean Intersection-Over-Union
(mIoU), Sørensen–Dice coeficient (DSC), recall, precision (Prec.),
accuracy (Acc.), and F2 metrics were used for both task 1 and 2.
Additionally, FPS was also calculated for task 2. Task 1 and task 2’s
evaluation score is the same as a single model was used for both the
tasks. The proposed model trained on the Kvasir-SEG dataset [
The Medico Automatic Polyp Segmentation challenge [