| Title |
Deep Learning Under Privileged Information for Pneumonia Detection |
| Authors |
고명섭(Myeongseob Ko) ; 정병창(ByeongChang Jeong) ; 김대겸(Daegyeom Kim) ; 한철(Cheol E. Han) |
| DOI |
https://doi.org/10.5573/ieie.2021.58.3.67 |
| Keywords |
deep learning; privileged information. pneumonia detection |
| Abstract |
In a recent computer vision society, there has been an rapid improvement in the performance of image classification tasks along with the development of deep learning technology. In the medical field, these classification techniques have been widely exploited to detect and diagnose several types of diseases. In this paper, we propose a method to detect pneumonia by additionally providing predetermined privileged information with off-the-shelf deep learning networks based on Learning Under Privileged Information (LUPI) framework. The privileged information is a designated area within an image, and can serve as implicit attention, encouraging the model to focus on the area directly related to the task, and thus may improve the classification performance. As an example, in this paper, we designated lung areas as our privileged information. Our proposed model consists of two shared VGG-16 models; one is for processing main information, image itself, and the other is for processing privileged information through information bottleneck, Gaussian dropout, and reparameterization trick. We provided various sized datasets but smaller than the original dataset by resampling it and compared model performances with and without privileged information. Our experiment showed that privileged information improves the test accuracy and F1 score, and the performance gain by the privileged information remarkably increases as the size of dataset gets smaller: increasing test accuracy and F1 score respectively by 3.5% and 0.0285 with 1000 training images, by 3% and 0.0173 with 100 training images, and by 16.72% and 0.0629 with 75 training images. We also demonstrated that our model can be interpretable through the activation maps of our model with the privileged information. |