| Resumo: | Retina image analysis is an important screening tool for early detection of multiple dis eases such as diabetic retinopathy which greatly impairs visual function. Image analy sis and pathology detection can be accomplished both by ophthalmologists and by the use of computer-aided diagnosis systems. Advancements in hardware technology led to more portable and less expensive imaging devices for medical image acquisition. This promotes large scale remote diagnosis by clinicians as well as the implementation of computer-aided diagnosis systems for local routine disease screening. However, lower cost equipment generally results in inferior quality images. This may jeopardize the reliability of the acquired images and thus hinder the overall performance of the diagnos tic tool. To solve this open challenge, we carried out an in-depth study on using different deep learning-based frameworks for improving retina image quality while maintaining the underlying morphological information for the diagnosis. Our results demonstrate that using a Cycle Generative Adversarial Network for unpaired image-to-image trans lation leads to successful transformations of retina images from a low- to a high-quality domain. The visual evidence of this improvement was quantitatively affirmed by the two proposed validation methods. The first used a retina image quality classifier to confirm a significant prediction label shift towards a quality enhance. On average, a 50% increase of images being classified as high-quality was verified. The second analysed the perfor mance modifications of a diabetic retinopathy detection algorithm upon being trained with the quality-improved images. The latter led to strong evidence that the proposed solution satisfies the requirement of maintaining the images’ original information for diagnosis, and that it assures a pathology-assessment more sensitive to the presence of pathological signs. These experimental results confirm the potential effectiveness of our solution in improving retina image quality for diagnosis. Along with the addressed con tributions, we analysed how the construction of the data sets representing the low-quality domain impacts the quality translation efficiency. Our findings suggest that by tackling the problem more selectively, that is, constructing data sets more homogeneous in terms of their image defects, we can obtain more accentuated quality transformations.
|
|---|