Deep Learning for Detection and Localization of Thoracic
Diseases using Chest X-Ray Imagery

Somnath Rakshit^1,2,+, Indrajit Saha^1,+,*, Michal Wlasnowolski^2,3, Ujjwal Maulik⁴, and Dariusz Plewczynski^2,3

¹Department of Computer Science and Engineering, National Institute of Technical Teachers' Training and Research, Kolkata, India
²Centre of New Technologies, University of Warsaw, Poland
³Faculty of Mathematics and Information Science, Warsaw University of Technology, Warsaw, Poland
⁴Department of Computer Science and Engineering, Jadavpur University, Kolkata, India
*Correspondence should be addressed to indrajit@nitttrkol.ac.in
⁺These authors contributed equally to this work

ABSTRACT

Classification of diseases from biomedical images is a fast growing emerging field of research. In this regard, chest X-Rays (CXR) are one of the most widely used medical images to diagnose common heart and lung diseases where previous works have explored the usage of various pre-trained deep learning models to perform the classification. However, these models are very deep, thus use large number of parameters. Moreover, it is still not possible to find readily available access to a practicing radiologist for proper diagnosis from an X-Ray image of chest. Hence, this fact motivated us to conduct this research with the aim to classify CXR images in an automated manner with smaller number of parameters during training for 14 different categories of thoracic diseases and produce heatmap for the corresponding image in order to show the location of abnormality. For the purpose of classification, transfer learning is used with the pre-trained network of Resnet18, while the heatmaps are generated using pooling along the channel dimension and then computing the average of class-wise features. The proposed model contains less parameters to train and provides better performance than the other models present in the literature. The trained model is then validated both quantitatively and visually by producing localized images in the form of heatmaps of the CXR images.