Identification of polyps in endoscopic images is critical for the diagnosis of colon cancer. Finding the exact shape and size of polyps requires the segmentation of endoscopic images. This research explores the advantage of using depth-wise separable convolution in the atrous convolution of the ResUNet++ architecture. Deep atrous spatial pyramid pooling was also implemented on the ResUNet++ architecture. The results show that architecture with separable convolution has a smaller size and fewer Giga-Floating Point Operations (GFLOPs) without degrading the performance too much.
Wireless capsule endoscopy (WSE) has been used for diagnosis for nearly 10 years now. WSE images provide diagnosis capability for many diseases such as colon cancer, ulcer, polyps detection, etc. With the advent of deep learning in computer vision, this diagnosis task can be automated.
The gastrointestinal tract has been an active area of research. The
benefit that can be achieved through computer-aided diagnosis
is significant. Jha et al. [
The approach follows the method used by Jha et al. [
Input Output
A bridge of pyramid pooling is used between encoder and
decoder block [
This Atrous Spatial Pyramid Pooling (ASPP) block in ResUNet++
was implemented using depth-wise separable convolution as well
as replaced with Deep Atrous Spatial Pyramid Pooling (DASPP)
module from [
Semantic segmentation, unlike object detection, can be treated
as a pixel wise classification problem. The output of semantic
segmentation for a pixel is a mask identifying the class to which the
pixel belongs. For the polyps segmentation problem [
The experiments were performed on Kvasir-SEG dataset [
All the experiments were performed on google Colab which provides a session for up to 12 hours. This 12 hours session is not enough to train a deep learning model. So to make a fair comparison, the number of epochs for all the experiments was kept the same. The data was split into training, validation, and testing set with the ratio of 80, 10, and 10 percent respectively. With this split, 800 images were selected for model training. These 800 images are not enough to train a deep learning model. To increase the training set data augmentation technique was applied to the training set. The validation set and testing set were not modified and thus the size of validation and test sets were 100 images each. 30 diferent augmentations were applied to the training set after that the size of it grew to 24800 images. The augmentations were also applied to the provided mask so that the target variable is transformed in the same way as the input image.
The optimizer used from training was NAdams optimizer with a learning rate of 0.0001 and a batch size of 8. The learning curve for training and validation loss was recorded for each epoch. The learning curve provides insights into the model convergence.
Figure 2 shows the learning curve for each architecture. The architecture with the DASPP bridge shows that it may have converged within 10 epochs as the validation error started increasing. However, the ResUNet++ and separable convolved ResUNet show that the model can be trained for few more epochs as both training and validation error are still decreasing. For UNet, the learning curve is also decreasing at the 10th epoch. However, the value of the loss is higher than the loss of ResUNet++ architecture.
Unet 75.23% 84.52% resunet++ 64.97% 89.81% sepv_conv_resunet++ 60.55% 93.31% dsapp_resunet++ 69.72% 82.62% dsapp_relu_resunet++ 61.54% 92.33% Table 1: Test Data Results
71.91% 78.35% 77.25% 76.66% 74.63% mIoU 59.53% 69.48% 67.56% 66.71% 66.03%
Table 1 gives the performance of each model on testing data. The performance of ResUNet++ on dice coeficient and mIoU is better than other models. The performance of the model with separable convolution has comparable results on Dice and mIoU metrics. However, the model with the DASPP bridge did not perform well. This shows that increasing depth any further did not improve performance. The size of the model which is measure by the number of parameters and Giga-Floating Point Operations (GFLOPs) is best for the model with separable convolution. The results are compiled in table 2. The less number of parameters means that the model size is smaller and it may be easy to move this in a production environment.
Unet 3,588,997 resunet++ 4,371,265 sepv_conv_resunet++ 3,047,265 dsapp_resunet++ 5,024,705 dsapp_relu_resunet++ 5,024,705 Table 2: Model Size
The research gives empirical results of the advantage of using depthwise separable convolution which resulted in smaller model size without significantly afecting the performance. It has also been shown that increasing the depth further may not improve performance and can result in overfitting of the model. It has been observed that the implementation of depth-wise separable convolution results in a smaller model without much degradation in overall performance. The tuning of hyper-parameter and a larger number of epochs will give a better understanding of the performance.