Deep learning based automated detection of intraretinal cystoid fluid

• Published in: International journal of imaging systems and technology 2022-05, Vol.32 (3), p.902-917
• Main Authors: Ahmed, Zeeshan, Panhwar, Shahbaz Qamar, Baqai, Attiya, Umrani, Fahim Aziz, Ahmed, Munawar, Khan, Arbaaz
• Format: Article
• Language: English
• Subjects: Algorithms; artificial intelligence (AI); Automation; cystoid macular edema (CME); Deep learning; deep learning (DL); diabetic macular edema (DME); Edema; Filtration; image processing; Image quality; Image segmentation; Machine learning; Model accuracy; Optical Coherence Tomography; optical coherence tomography (OCT); Performance evaluation; Performance measurement; Quality assessment; Vision
• ISSN: 0899-9457; 1098-1098
• DOI: 10.1002/ima.22662

The diversified ocular disorders in which cystoid macular edema (CME) occurs, are strongly associated with the vision loss. Optical coherence tomography (OCT) scans that allow screening of the retina, contain artifacts including blur‐edges, speckle noise, and so forth, which create difficulty in identifying retinal fluid. In this work, major image preprocessing techniques such as minimum filtering, block‐matching and 3D filtering, and Richardson–Lucy deconvolution method are applied to minimize noise and other degradation effects from OCT scans, ensured by maintaining image quality assessment scores, that is, blind‐less image spatial quality evaluator and sharpness estimation score. Furthermore, this work develops an automated method based on deep learning (DL) to detect the presence and progression of retinal fluid, that is, CME and its quantification. Specifically, U‐net model with efficient tuning of hyperparameters is proposed to obtain optimal results. The algorithm's performance is evaluated for accurate identification of fluid localization in the case of diabetic macular edema (DME), having CME regions, and compared against the manual segmentation of fluid by experts, which are highly correlated. This enhanced DL method achieves better performance than the algorithms previously reported. The proposed algorithm is evaluated on several performance metrics during training and reports the best score of 99.81, 0.50, 83.34, 86.56, 80.89, and 99.48% for model accuracy, binary cross‐entropy loss, Dice coefficient (DC), precision, recall, and area under the ROC‐curve, respectively. Additionally, fourfold cross‐validation technique is also applied and an average DC score of 84.00% is obtained. This approach can be helpful to prevent vision loss.