Week 6— Detecting Musculoskeletal Conditions
Team members: Utku İpek, Hüseyincan Kaynak, Gokce Sengun
Our previous blog post: Week-5 Detecting Musculoskeletal Conditions
Hello everyone!
This is our latest blog on Musculoskeletal Conditions. Today we will talk to you about the results and achievements of our project. Let’s start!
When we started our project we used our Convolutional Neural Network (CNN) model. And we have made many improvements to improve this model. Then we used Densely Connected Convolutional Networks (DenseNet) as an alternative to CNN. If you want to know more about CNN and DenseNet, please see our previous blog.
We selected “wrist” and “hand” from the sections in our data set to use in these two models. We will first show photos of the improvements in our CNN model for the photos we used for testing, then we will share the accuracy and loss values of our data sets in these two models and then share our analysis on these results.
Data Augmentation Images
The left image shows that we use different zooming features to improve our CNN model. The right image shows that we use these images in different rotations.
Loss and Accuracy for CNN
We got the results in our CNN model for Wrist. When the results were examined, although the training of the model seems to be unfinished, the graphic continued in this way and we cut it to 100 epochs because there was no change.
Loss and Accuracy for DenseNet
We used the wrist region for our DenseNet result. When we look at the figure in the loss graph, we see that after 10 epochs our model started to memorize. Our DenseNet results are better than those of our CNN model.
Confusion Matrix for CNN
The results in the CNN model are visualized as Confusion Matrix. The left side is the normalized state, the right side is the normalized state.
The test result of our model is 70%.
Confusion Matrix for DenseNet
The results in the DenseNet model are visualized as Confusion Matrix.In this picture, the results of the DenseNet model are visualized as Confusion Matrix. The left side is the normalized state, the right side is the normalized state.
The test result of our model is 84%.
When we look at the two results, we see that DenseNet works better for a small number of data. We were unlucky in our test results because we didn’t have our test data and we had to split our train data for the test data. When we look at daily life, the results of our models are also very good because the success rates of real doctors are not very high.
Our next goal is to make a web application part of this project. In this way, it is possible for most people to detect the abnormalities of images downloaded from the internet. In this web application, heat map visualization also makes it easier to detect areas with abnormalities.
You deserve a big applause for sharing the excitement of working with us for weeks! Don’t forget to follow us for our next works. You can watch the video of our project here. Enjoy the video :)Good Bye!
