Week 5— Detecting Musculoskeletal Conditions
Team members: Utku İpek, Hüseyincan Kaynak , Gokce Sengun
Our previous blog post: Week-4 Detecting Musculoskeletal Conditions
Hello! Last week we created a model with CNN, we tried the humerus bone images from the dataset and we got 65% accuracy for the first trial. This result is not bad for the model we created, but it may be even better.
Because we created a model without considering Image Data Augmentation. Therefore, the Image Data Augmentation we use is not enough for the model accuracy. So what is the concept of this Image Data Augmentation? Let’s look at it all together. This week we will tell you how to improve CNN performance.
There are many methods to improve CNN performance. These methods are as follows.
1. Tune Parameters,
2. Handel Overfitting and Underfitting problem,
3. Deeper Network Topology,
4. Image Data Augmentation.
After briefly discussing these methods, we will discuss Image Data Augmentation in this blog. We will see how this improvement can benefit our model for separate parts of the body in our data set.
1.Tune Parameters:
To improve CNN model performance, we can tune parameters like epochs, learning rates, etc. The number of epochs we have identified during the training phase of our model is important to improve performance. After choosing the right epoch, we can see that there is no reduction, loss of training and improvement in training accuracy.
2. Handle Overfitting and Underfitting problem:
It is very important to use the data correctly when training a model. If the data is not used correctly, problems such as overfitting and underfitting may occur.
In the overfitting, your model gives very nice accuracy on trained data but very little accuracy on test data. The meaning of this is overfitting model is having good memorization ability but less generalization ability. Our model doesn’t generalize well from our training data to unseen data.
Underfitting refers to a model that works well on the testing data. Which means we don’t know enough about our data. So we can’t analyze the data enough. As a solution, we need to get more data features.
3. Deeper Network Topology :
A wide neural network is possible to train with every possible input value. Hence, these networks are very good at memorization, but not so good at generalization. There are, however, a few difficulties with using an extremely wide, shallow network. These neural networks are very costly in practical applications.
We want our network to be as small as possible to produce good results. The wider network will take a longer time to train. Deep networks are very computationally expensive to train. Hence, make them wide and deep enough that they work well, but no wider and deeper.
4. Image Data Augmentation :
The performance of deep learning neural networks often improves with the amount of data available. If we have not enough data our model can’t predict accurately. If you are such as position like not have a huge amount of data solution is data augmentation.
Data Augmentation is a technique that creates new training data with your existing data. While doing that data augmentation use some operators like; shifts , zooms, rotates, flips, etc. The main purpose here is that expand our existing training data. In the test stage, may the image be taken from different angles or upside down in such cases data augmentation can handle.
Our Implementation Result
In the first image, we got a result without Image Data Augmentation. According to this result in the last epoch;
Train Accuracy: 0.9921
Train Loss: 0.0426
Validation Accuracy: 0.5729
Validation Loss: 5.3095
In the second image, we got a result with Image Data Augmentation. According to this result in the last epoch;
Train Accuracy: 0.6411
Train Loss: 0.6288
Validation Accuracy: 0.5602
Validation Loss: 0.8426
Take a look at the above graphics are one of them with data augmentation and the other one is not. Our test part of the body which is the humerus, we don’t have a huge amount of image. That’s why we used data augmentation and we want to compare results with what’s happened if we don’t use it. We can obtain while looking at these graphics Data Augmentation is a significant effect our error. Because we expand our train data examples. If you stuck on the fewer example you can try data augmentation.
This week we talk about improving CNN performance, what are the methods to improve and how we do that. While you reading these lines we will enter a new era. 2020 is at our door. We wish you all a happy and productive new year. We will see you next week, we are getting closer to the end of the project.
