Quality Evaluation Of Medical Images Based On Deep Learning

Precise diagnosis and treatment depend on high-quality medical images,which can only be acquired with adequate image quality evaluation and assurance.At present,image quality assessment of the radiology department mainly depends on the subjective evaluation by senior radiologists.This is not only easily influenced by the subjectivity of the radiologists,but also places an extra burden on them.Therefore,an effective,objective,automatic image quality evaluation system is always in need in hospitals.In recent years,deep learning has made great progresses in the field of computer vision,such as object detection and classification.Inspired by these achievements,we used convolutional neural networks to study two problems in medical image quality evaluation: one is the nasopharyngeal image quality control in magnetic resonance imaging(MRI)images;the other is the quality control of chest digital radiography(DR)images.We proposed a method for automatic image quality assessment of the nasopharyngeal region in MRI images.A YOLOv3 network was trained to detect the nasopharyngeal region in MRI images and a convolutional neural network(CNN)based on the residual network and the squeeze excitation layer was used to classify the qualities of the region of interest(ROI)extracted from the MRI.The experimental results demonstrated that the proposed algorithm can be used to identify images of poor quality.The quality of chest DR images is affected by many factors,such as the posture of the patient,foreign objects,equipment parameters,etc.For automatic quality assessment of chest DR images,we proposed two solutions: a two-step patch-based CNN,and a multi-label classification network.The experimental results showed that both models can be used to assess the quality of chest DR images.We also integrated the trained models into a home-made software to facilitate their use in the clinical environment.With appropriate authorization and configuration,the software can automatically access the newly acquired chest DR images and perform the quality assessment online.This can help to increase the ratio of high-quality chest DR images and improve the image-based clinical diagnosis.