Page 118 - Computational Retinal Image Analysis

5 Summary and outlook 111

Moreover, Guimaraes et al. [147] report segmentation and reconstruction of 3D
retinal vasculature. With the further advancement of 3D imaging techniques,
this exciting direction of 3D vessel analysis possesses excellent research
potential. It is also well connected to existing 3D blood vessel analysis efforts of
other organs and via different instruments such as magnetic resonance imaging.
Multimodal vessel analysis. One prominent example of multimodal retinal
image analysis could be the combined usage of 2D fundus and 3D OCT
images in vessel segmentation and tracing. Here, Hu et al. [148] investigate
the integration of 2D fundus image and corresponding 3D OCT images in
retinal vessel segmentation. SLO fundus image and corresponding macular
OCT slices are jointly considered in Ref. [149] in two steps. The first step
involves 2D vessel segmentation of fundus images in curvelet domain, with
the side information from multiple OCT slices. The second step focuses on 3D
reconstruction of the blood vessels from the OCT data. In addition to images,
patient-related information such as electronic medical records and genetic
data could also be taken into account for better disease diagnosis, and patient
prognosis and treatment.
Analyzing mobile retinal images. As stated in Ref. [150], the recent development
of portable fundus cameras [151] and smartphone-based fundus imaging
systems have led to considerable opportunities as well as new challenges to
retinal image analysis, such as the demand for more affordable imaging devices
with perhaps lower computation cost [152]. To address the need of the emerging
mobile retinal diagnostic devices that calls for segmentation techniques with
low memory and computation cost, Hajabdollahi et al. [152] propose to train
a simple CNN model-based pruning and quantization of the original full-
sized model, that is capable of retaining the performance with much reduced
computation and size. Looking forward, we expect further and considerable
progresses along this direction.

5.2 Benchmarks and metrics
Benchmark datasets, such as ImageNet [39] and COCO [40], have played a vital role
in fueling the recent computer vision breakthroughs. It has been observed that these
successful datasets are both of large scale, and richly annotated. For example, over
the years, COCO has introduced and aggregated various annotations on categories
and shapes of individual objects and stuffs (backgrounds). Currently, it is capable
of hosting a broad variety of closely related tasks including instance segmentation,
stuff segmentation, object detection, person keypoint localization, as well as image
captioning, visual dialog, image attributes, text detection and recognition, among
others. By contrast, existing retinal vasculature image datasets, as shown in Table 1,
are of small size, and often not richly annotated. While most state-of-the-art results
are still reported based on the well-known benchmarks of DRIVE and STARE, the
low-resolution images considered in these datasets are considerably lagging behind