Application And Evaluation Of Deep Learning-based Image Enhancement Techniques In Optical Coherence Tomography

Purpose:To develop a deep learning-based framework combined with High-Resolution Network(HRNet)to improve the image quality of optical coherence tomography(OCT)and evaluate its image enhancement as well as automatic segmentation effect with traditional image averaging method from a clinical perspective.Methods:In this study,a deep learning architecture for OCT image enhancement is developed.It was constructed based on the idea of U-net with a high-resolution representation block.In particular,the network inputs were designed as multiple OCT frames in order to acquire more original scanned information during image enhancement,and the number of network input ranged from 1 to 20.We recruited 196 healthy eyes and 251 diseased eyes with various retinal diseases.460 sets(200 healthy eyes and 260 diseased eyes)from 610 sets were used as the training dataset,and the rest 150 sets(64 healthy eyes and 86 diseased eyes)were used as validation dataset.Quantitative assessment was performed from objective and subjective perspectives.Three objective metrics were applied including structural similarity index measure(SSIM),peak signal-to-noise ratio(PSNR)and contrast-to-noise ratio(CNR)of regions of interest(ROI)were performed between deep learning method and traditional image averaging.Subjective assessment was performed by four ophthalmologists on 205 image sets.We calculated SSIM,PSNR and CNR between standard image and each enhanced image by different methods.Then paired sample t-tests were applied to compare the subjective scores between each pair of enhanced images for retinal anatomical structures and different retinal lesions,respectively.Results:This study revealed that with the increase of frame count from 1 to 20,SSIM values were increased from 0.83 to 0.97 for deep learning method,and from 0.47 to 0.95 for traditional averaging.Meanwhile,PSNR values were increased from 30.23 to 40.1 for deep learning method,and from 23.93 to 39.4 for averaging method.The result indicating that the deep learning algorithm obtained better effect than the traditional averaging in improving OCT image quality and deep learning method shows more significant enhancement ability when the frame count of the input image is small.We observed that the deep learning method with 5 frames could achieve the comparable SSIM with averaging method with 16 frames,therefore we selected 5 frames as the inputs of separate image-enhancement method,the local objective assessment with CNR of ROI illustrated that the deep learning method had more obvious tissue contrast enhancement than averaging method.The result revealed that the enhanced image produced by the deep learning framework has higher contrast than the traditional image averaging and the enhanced image produced by the deep learning framework has more similar visual effects with standard image,including less speckle noise,better contrast and smoother edges.The subjective scores of image quality were all higher in our deep learning method than in traditional averaging method,not only for normal retinal structure but also for various retinal lesions.And all the objective and subjective indicators showed significant statistically differences(P<0.05).Conclusion:Compared to traditional image averaging,our proposed framework based on deep learning architecture can achieve a reasonable trade-off between image quality and scanning times,with the benefit of the reduced number of repeated scans.The outcomes of image enhancement were evaluated both objectively and subjectively on OCT images acquired from healthy individuals and patients with retinal diseases.