MIDC-Net: Medical Image Denoising and Disease Classification Network for Chest X-rays

Accurate medical imaging is vital for precise disease diagnosis and effective treatment. However, X-ray images may be subject to varying degrees of noise due to factors such as patient health conditions requiring reduced X-ray radiation. In this paper, the Medical Image Denoising and Disease Classif...

Full description

Saved in:
Bibliographic Details
Published in2023 3rd International Conference on Electronic Information Engineering and Computer Science (EIECS) pp. 834 - 838
Main Author Li, Jiatu
Format Conference Proceeding
LanguageEnglish
Published IEEE 22.09.2023
Subjects
convolutional neural network
Decoding
denoising autoencoder
Diseases
Image denoising
Image reconstruction
multi-label classification
Noise measurement
Noise reduction
X-ray imaging
Online AccessGet full text

Cover

More Information
Summary:Accurate medical imaging is vital for precise disease diagnosis and effective treatment. However, X-ray images may be subject to varying degrees of noise due to factors such as patient health conditions requiring reduced X-ray radiation. In this paper, the Medical Image Denoising and Disease Classification Network (MIDC-Net) is proposed to solve the issue of noisy chest X-rays that negatively affects thorax disease classification by performing image denoising and disease classification simultaneously. The MIDC-Net integrates a Convolutional Neural Network (CNN) for classification, and a denoising autoencoder with convolutional layers and residual connections for image denoising. It comprises three components, namely an encoder, a decoder, and a classifier, where the input of the encoder is noisy images with varying levels of Gaussian noise, and the output of the decoder and the classifier are reconstructed images and predicted diseases. The loss of the MIDC-Net is a linear combination of Mean Square Error (MSE) loss used for denoising and Binary Cross-Entropy (BCE) loss used for multi-label classification. Experimental outcomes on the NIH chest X-ray dataset illustrated that the reconstructed images resembled the original X-rays, and the network achieved a low MSE loss, but the quality of denoised images exhibited an inverse relationship with noise levels. Regarding multi-label thorax classification results, the test accuracies all exceed 95% regardless of noise levels of the input images, demonstrating efficient and accurate prediction in the scenario of noisy inputs.
DOI:10.1109/EIECS59936.2023.10435558