The Medico: Multimedia Task 2020 focuses on developing an eficient and accurate computer-aided diagnosis system for automatic segmentation [3]. We participate in task 1, polyps segmentation task, which is to develop algorithms for segmenting polyps on a comprehensive dataset. In this task, we propose methods combining Residual module, Inception module, Adaptive Convolutional neural network with U-Net model and PraNet for semantic segmentation of various types of polyps in endoscopic images. We select 5 runs with diferent architecture and parameters in our methods. Our methods show potential results in accuracy and eficiency through multiple experiments, and our team is in Top 4 best results with Jaccard index of 0.765.
The goal of Medico automatic polyp segmentation challenge is to evaluate various methods for automatic polyp segmentation that can be used to detect and mask out various types of polyps (including irregular, small or flat polyps) with high accuracy.
In this challenge, our goals are to segment the mask of all types
of polyps in the dataset. We consider five solutions, corresponding
to our five submitted runs. In our first approach, we adopt a simple
U-Net architecture [
We submit five runs for this task. In each run, we modify and improve diferent architectures to enhance accuracy and segmentation speed. Initially, we choose U-Net to create the base model that we can test the baseline of the test set. However, the evaluation get low, we decide to develop the U-Net by using Leaky ReLU, but the result gets lower. Next, we replace the simple Convolution block with a more complicated block, the Inception block and Residual block, to improve the accuracy. However, the results are not higher. So, we decide to choose PraNet and get the results as our expectation. 2.1
In the first run, to have the overview of the segmentation problem
in this task, we use U-Net, which is used widely in medical image
segmentation [
In our model, we keep the encoding step - which includes 5x5 and 3x3 convolutional blocks and a 1x1 convolutional block with ReLU. However, in this run, to fit the model, we use only one Pooling layer with a small learning rate and batch size. 2.2
In the second run, to enhance the result with U-Net, we breakdown and modify the ReLU layer. During the training step, ReLU can cause something known as dead neurons. To overcome this critical defect of dying neurons, we substitute ReLU with another activation function, the Leaky ReLU. We replace all ReLU activation functions with our custom leaky ReLU blocks while preserving the remaining convolutional blocks and layers. Moreover, after each convolutional block, we also add a Leaky ReLU to evaluate extracted features. 2.3
In the third run, we propose to extract more features to improve the
model. We add a more standard convolutional block to this
structure, but it easily to be a vanishing gradient. To achieve consistent
training as the depth of the network increases and also prevent
those issues, we replace the building blocks of the U-Net
architecture with modified residual blocks of convolutional layers. [
We combine the Inception module with U-Net. The main idea of the Inception architecture based on finding out how an optimal local sparse structure in a convolutional vision network can be approximated and covered by readily available dense components.
One of the main beneficial aspects of this architecture is that it allows for increasing the number of units at each stage significantly without an uncontrolled blow-up in computational complexity. The ubiquitous use of dimension reduction allows for shielding a large number of input filters of the last step to the next layer, first reducing their dimension before convolving over them with large patch size.
Another practically useful aspect of this design is that it aligns
with the intuition that visual information should be processed at
various scales and then aggregated so that the next stage can
abstract features from diferent scales simultaneously. The eficient
allocation of computational resources allows both the width of
each step and the total number of steps to be increased without
getting computational dificulty. [
In the last run, we use PraNet, which is a parallel reverse attention
network for accurate polyp segmentation in colonoscopy and also
has the highest benchmark on Medical Image Segmentation on
Kvasir-SEG [
The Kvaris-SEG [
Data preparation: We use Retina net to create a bounding box
cover the Polyps in endoscopic images.[
Data Augmentation: To increase the numbers of data to improve the training process, we use data augmentation with two methods: rotation and zooming. 3.2
Evaluation is on the test set of Mediaeval- Medico task, which include 160 images of Polyps in endoscopic images. Table 1 shows that Run 1 is better than Run 2 and Run 3 is better than Run 4. Run 5 is the best model overall. By replace ReLU with Leaky ReLU, although Accuray in Run 2 is not better than Run 1, Recall is gradually higher. Moreover, in 5 runs, ResUNet obtains the result of 90.1, which is approximate with the highest overall is 94.6.
Run ID Run 1 Run 2 Run 3 Run 4 Run5
Currently, there is a growing interest in the development of computeraided diagnosis (CADx) systems that could act as a second observer and digital assistant for the endoscopists. Algorithmic benchmarking is an eficient approach to analyze. We can try more new method with a diferent approach and improve the data preparation process to get a higher result.
In general, PraNet still gets the highest accuracy score. However, The computing speed remains high while U-Net and ResUNet take more advantages in speed with slightly diferent in results. Moreover, data preparation is the most impact process to achieve higher accuracy and evaluation metrics. Therefore, we propose to use Inception Resnet or Densenet 169 feature extractor combine with U-Net to improve the accuracy of the test.
This research is funded by Vietnam National University HoChiMinh City (VNU-HCM) under grant number D2020-42-01 on “Artificial Intelligence and Extended Reality for Medical Diagnosis and Treatment Assistance”.