11 Segmentation and Visualization of Drusen …
303
Also the RSVP method was efficient for different types and morphologies of
drusen. Due to the pixel filling step in the RSVP method, drusen with low or medium
reflectivity (such as the drusen marked with the blue triangle in Fig. 11.13), was able
to be visualized. The dark-region filling step had less influence on the visibility on
drusen with high reflectivity (such as the drusen marked with the green triangle in
Fig. 11.13 in the RSVP images since they were already highly visible by virtue of
their inherent pixel brightness and high reflectivity. Alternatively, all of the drusen
present were able to be visualized in a single RSVP image, thereby allowing for
easier identification of drusen. Some related work has been done in producing a
“slab” SVP of the retina to improve visualization of drusen (Cirrus SD-OCT, Carl
Zeiss Meditec, Inc, Software version 6.0.1) which is similar with the Georczynska’s
method [43]. To our knowledge there have been no articles on this method published
in the peer-reviewed literature; however, based on our understanding of the method
in the Cirrus system, the RSVP method is different, new and novel in that (1) it is
fully automated (the Cirrus software requires input from the operator to specify the
“slab” to be processed), (2) the “slab” method includes the retina from the RNFL
to the RPE, whereas the RSVP method includes a highly restricted volume of the
OCT scan in proximity to the RPE where drusen reside; thus, one expects that
drusen visualization in the RSVP approach will be degraded less than with the “slab”
method, and (3) the RSVP method incorporates image processing to enhance the
conspicuity of drusen by filling in the dark regions within drusen with bright pixels.
With respect to the qualitative and quantitative analysis above, the RSVP method is
more effective for drusen visualization than the SVP method, and more convenient
than the manual marking method [37] and the selective depth level method [38] which
can be important for ophthalmologists to directly and rapidly assess the macula of
patients with non-exudative age-related macular degeneration. Our future work will
be to undertake a comparison of the RSVP method to the manual “slab” method.
Some of the limitations of the present study will be discuss below. Because the filling of the dark region beneath drusen and extraction of the RPE layer and are based on
automatic algorithms, errors may occur in the RPE-based fundus image. For instance,
the red circled regions in Fig. 11.36 show areas which are not actual drusen, but rather
areas caused by the incorrect extraction of the RPE layer. Till now, these errors are
minimized or constitute to no affect on drusen visualization. Another limitation is
that filling the region under drusen with bright pixels may inadvertently fill (and
therefore obscure) small, focal pigment epithelial detachments (PEDs), incorrectly
characterizing them as drusen. Although this did not occur during this research, the
automated algorithm used will be modified in future work to recognize the relatively
dark areas under PEDs (as opposed to brighter areas contained in drusen) and exclude
such regions from the filling-in process, thereby avoiding this potential problem. The
last limitation was that the overlap ratio with the gold standard was imperfect, being
derived from two readers. An accurate gold standard would be based on histological
slides, which would be very difficult to obtain. More so, the quantity of scans and
readers we included in our quantitative evaluation was relatively small. Since each
reader needed to review many individual B-scans for drusen in each case, it would
have been too labor intensive and time consuming to manually segment many cases.
303
Also the RSVP method was efficient for different types and morphologies of
drusen. Due to the pixel filling step in the RSVP method, drusen with low or medium
reflectivity (such as the drusen marked with the blue triangle in Fig. 11.13), was able
to be visualized. The dark-region filling step had less influence on the visibility on
drusen with high reflectivity (such as the drusen marked with the green triangle in
Fig. 11.13 in the RSVP images since they were already highly visible by virtue of
their inherent pixel brightness and high reflectivity. Alternatively, all of the drusen
present were able to be visualized in a single RSVP image, thereby allowing for
easier identification of drusen. Some related work has been done in producing a
“slab” SVP of the retina to improve visualization of drusen (Cirrus SD-OCT, Carl
Zeiss Meditec, Inc, Software version 6.0.1) which is similar with the Georczynska’s
method [43]. To our knowledge there have been no articles on this method published
in the peer-reviewed literature; however, based on our understanding of the method
in the Cirrus system, the RSVP method is different, new and novel in that (1) it is
fully automated (the Cirrus software requires input from the operator to specify the
“slab” to be processed), (2) the “slab” method includes the retina from the RNFL
to the RPE, whereas the RSVP method includes a highly restricted volume of the
OCT scan in proximity to the RPE where drusen reside; thus, one expects that
drusen visualization in the RSVP approach will be degraded less than with the “slab”
method, and (3) the RSVP method incorporates image processing to enhance the
conspicuity of drusen by filling in the dark regions within drusen with bright pixels.
With respect to the qualitative and quantitative analysis above, the RSVP method is
more effective for drusen visualization than the SVP method, and more convenient
than the manual marking method [37] and the selective depth level method [38] which
can be important for ophthalmologists to directly and rapidly assess the macula of
patients with non-exudative age-related macular degeneration. Our future work will
be to undertake a comparison of the RSVP method to the manual “slab” method.
Some of the limitations of the present study will be discuss below. Because the filling of the dark region beneath drusen and extraction of the RPE layer and are based on
automatic algorithms, errors may occur in the RPE-based fundus image. For instance,
the red circled regions in Fig. 11.36 show areas which are not actual drusen, but rather
areas caused by the incorrect extraction of the RPE layer. Till now, these errors are
minimized or constitute to no affect on drusen visualization. Another limitation is
that filling the region under drusen with bright pixels may inadvertently fill (and
therefore obscure) small, focal pigment epithelial detachments (PEDs), incorrectly
characterizing them as drusen. Although this did not occur during this research, the
automated algorithm used will be modified in future work to recognize the relatively
dark areas under PEDs (as opposed to brighter areas contained in drusen) and exclude
such regions from the filling-in process, thereby avoiding this potential problem. The
last limitation was that the overlap ratio with the gold standard was imperfect, being
derived from two readers. An accurate gold standard would be based on histological
slides, which would be very difficult to obtain. More so, the quantity of scans and
readers we included in our quantitative evaluation was relatively small. Since each
reader needed to review many individual B-scans for drusen in each case, it would
have been too labor intensive and time consuming to manually segment many cases.