Limits...
2D Fast Vessel Visualization Using a Vessel Wall Mask Guiding Fine Vessel Detection.

Raptis S, Koutsouris D - Int J Biomed Imaging (2010)

Bottom Line: Sparse pixels are effectively eliminated by applying a limited range Hough Transform (HT) or region growing.Major benefits are limiting the range of parameters, reducing the search-space for post-convolution to only masked regions, representing almost 2% of the 2D volume, good speed versus accuracy/time trade-off.Results show the potentials of our approach in terms of time for detection ROC analysis and accuracy of vessel pixel (VP) detection.

View Article: PubMed Central - PubMed

Affiliation: Biomedical Engineering Laboratory, School of Electrical & Computer Engineering, National Technical University of Athens, 9 Iroon Polytechniou Str., H/Y Building-Zografou Campus, 15773 Athens, Greece.

ABSTRACT
The paper addresses the fine retinal-vessel's detection issue that is faced in diagnostic applications and aims at assisting in better recognizing fine vessel anomalies in 2D. Our innovation relies in separating key visual features vessels exhibit in order to make the diagnosis of eventual retinopathologies easier to detect. This allows focusing on vessel segments which present fine changes detectable at different sampling scales. We advocate that these changes can be addressed as subsequent stages of the same vessel detection procedure. We first carry out an initial estimate of the basic vessel-wall's network, define the main wall-body, and then try to approach the ridges and branches of the vasculature's using fine detection. Fine vessel screening looks into local structural inconsistencies in vessels properties, into noise, or into not expected intensity variations observed inside pre-known vessel-body areas. The vessels are first modelled sufficiently but not precisely by their walls with a tubular model-structure that is the result of an initial segmentation. This provides a chart of likely Vessel Wall Pixels (VWPs) yielding a form of a likelihood vessel map mainly based on gradient filter's intensity and spatial arrangement parameters (e.g., linear consistency). Specific vessel parameters (centerline, width, location, fall-away rate, main orientation) are post-computed by convolving the image with a set of pre-tuned spatial filters called Matched Filters (MFs). These are easily computed as Gaussian-like 2D forms that use a limited range sub-optimal parameters adjusted to the dominant vessel characteristics obtained by Spatial Grey Level Difference statistics limiting the range of search into vessel widths of 16, 32, and 64 pixels. Sparse pixels are effectively eliminated by applying a limited range Hough Transform (HT) or region growing. Major benefits are limiting the range of parameters, reducing the search-space for post-convolution to only masked regions, representing almost 2% of the 2D volume, good speed versus accuracy/time trade-off. Results show the potentials of our approach in terms of time for detection ROC analysis and accuracy of vessel pixel (VP) detection.

No MeSH data available.


Related in: MedlinePlus

(a) Original fundus image with not well-traced peripheral network. (b) Original fundus image ellipse corrected. (c) Binary vessel wall pixel map (VWP). (d) Vessel map based on VWP. (e) Retinal Vasculature based on VWP and fine segmented for local details. (f) Retinal Vasculature without VWP map using MIP. (g) Ground troth vessel vasculature. (h) Redundant vessel vasculature produced using an Augmented VM and double-sided GS-MG. (i) Successive maps taken using different features and different sensitivity. (A) Gradient homogeneity VMs. (B) Paired VP map and dominant gradient VM. (C) Increased vessel width allowed, lower gradient threshold and interpolated WPs. (D) Experimentally proved VM.
© Copyright Policy - open-access
Related In: Results  -  Collection


getmorefigures.php?uid=PMC2913788&req=5

fig1: (a) Original fundus image with not well-traced peripheral network. (b) Original fundus image ellipse corrected. (c) Binary vessel wall pixel map (VWP). (d) Vessel map based on VWP. (e) Retinal Vasculature based on VWP and fine segmented for local details. (f) Retinal Vasculature without VWP map using MIP. (g) Ground troth vessel vasculature. (h) Redundant vessel vasculature produced using an Augmented VM and double-sided GS-MG. (i) Successive maps taken using different features and different sensitivity. (A) Gradient homogeneity VMs. (B) Paired VP map and dominant gradient VM. (C) Increased vessel width allowed, lower gradient threshold and interpolated WPs. (D) Experimentally proved VM.

Mentions: For the sake of performance comparison we give in Figures 1(c) and 1(e) the results using our VWM-based detection algorithm and the manual map produced as in [10]. The manual map is very similar to the detection map produced in the method given in [10]. However, using our structure-sensitive MF as in [1], we have a much finer structure not illustratable neither on the manual map nor in the computer detection map. The obvious but shallow drawback of our method is that we achieve at worse 10% higher number of FP's at the expense of speed. The gold standard map is not as clear though to easily visualize as the small vessel endings or patterns. Still, our result is largely justified by the additional argument that it is better to have more information than less and at better speed as seen in our results for single and dual mode GS-MF. Especially in our case the essential fine and coarse vasculature information is there. This eliminates the need for tedious metaprocessing as would be the case if we had a proven large number of FP's and the ground truth (GT) was in direct incoherence to our detection results. We also argue that the exemplary maps given in the literature are very good but still are estimates of the vessel network. Hence, there is nothing to prove in an absolute and not arguable manner that any possible abnormality can be studied using these GT binary maps. In fact, as it is sufficiently documented and mathematically outlined in [11, 12], we need to combine a capable number of human observers' binary maps to have an acceptably supported GT map as a gold standard. Then we also need to provide the GT results for all logical operations performed in each binary observation map separately. To this we need to apply special statistics as to how one can pool out with confidence experts' binary observations and with which confidence. These binary experts from many relatively and not absolutely reliable observers need to establish a more reliable standard.


2D Fast Vessel Visualization Using a Vessel Wall Mask Guiding Fine Vessel Detection.

Raptis S, Koutsouris D - Int J Biomed Imaging (2010)

(a) Original fundus image with not well-traced peripheral network. (b) Original fundus image ellipse corrected. (c) Binary vessel wall pixel map (VWP). (d) Vessel map based on VWP. (e) Retinal Vasculature based on VWP and fine segmented for local details. (f) Retinal Vasculature without VWP map using MIP. (g) Ground troth vessel vasculature. (h) Redundant vessel vasculature produced using an Augmented VM and double-sided GS-MG. (i) Successive maps taken using different features and different sensitivity. (A) Gradient homogeneity VMs. (B) Paired VP map and dominant gradient VM. (C) Increased vessel width allowed, lower gradient threshold and interpolated WPs. (D) Experimentally proved VM.
© Copyright Policy - open-access
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC2913788&req=5

fig1: (a) Original fundus image with not well-traced peripheral network. (b) Original fundus image ellipse corrected. (c) Binary vessel wall pixel map (VWP). (d) Vessel map based on VWP. (e) Retinal Vasculature based on VWP and fine segmented for local details. (f) Retinal Vasculature without VWP map using MIP. (g) Ground troth vessel vasculature. (h) Redundant vessel vasculature produced using an Augmented VM and double-sided GS-MG. (i) Successive maps taken using different features and different sensitivity. (A) Gradient homogeneity VMs. (B) Paired VP map and dominant gradient VM. (C) Increased vessel width allowed, lower gradient threshold and interpolated WPs. (D) Experimentally proved VM.
Mentions: For the sake of performance comparison we give in Figures 1(c) and 1(e) the results using our VWM-based detection algorithm and the manual map produced as in [10]. The manual map is very similar to the detection map produced in the method given in [10]. However, using our structure-sensitive MF as in [1], we have a much finer structure not illustratable neither on the manual map nor in the computer detection map. The obvious but shallow drawback of our method is that we achieve at worse 10% higher number of FP's at the expense of speed. The gold standard map is not as clear though to easily visualize as the small vessel endings or patterns. Still, our result is largely justified by the additional argument that it is better to have more information than less and at better speed as seen in our results for single and dual mode GS-MF. Especially in our case the essential fine and coarse vasculature information is there. This eliminates the need for tedious metaprocessing as would be the case if we had a proven large number of FP's and the ground truth (GT) was in direct incoherence to our detection results. We also argue that the exemplary maps given in the literature are very good but still are estimates of the vessel network. Hence, there is nothing to prove in an absolute and not arguable manner that any possible abnormality can be studied using these GT binary maps. In fact, as it is sufficiently documented and mathematically outlined in [11, 12], we need to combine a capable number of human observers' binary maps to have an acceptably supported GT map as a gold standard. Then we also need to provide the GT results for all logical operations performed in each binary observation map separately. To this we need to apply special statistics as to how one can pool out with confidence experts' binary observations and with which confidence. These binary experts from many relatively and not absolutely reliable observers need to establish a more reliable standard.

Bottom Line: Sparse pixels are effectively eliminated by applying a limited range Hough Transform (HT) or region growing.Major benefits are limiting the range of parameters, reducing the search-space for post-convolution to only masked regions, representing almost 2% of the 2D volume, good speed versus accuracy/time trade-off.Results show the potentials of our approach in terms of time for detection ROC analysis and accuracy of vessel pixel (VP) detection.

View Article: PubMed Central - PubMed

Affiliation: Biomedical Engineering Laboratory, School of Electrical & Computer Engineering, National Technical University of Athens, 9 Iroon Polytechniou Str., H/Y Building-Zografou Campus, 15773 Athens, Greece.

ABSTRACT
The paper addresses the fine retinal-vessel's detection issue that is faced in diagnostic applications and aims at assisting in better recognizing fine vessel anomalies in 2D. Our innovation relies in separating key visual features vessels exhibit in order to make the diagnosis of eventual retinopathologies easier to detect. This allows focusing on vessel segments which present fine changes detectable at different sampling scales. We advocate that these changes can be addressed as subsequent stages of the same vessel detection procedure. We first carry out an initial estimate of the basic vessel-wall's network, define the main wall-body, and then try to approach the ridges and branches of the vasculature's using fine detection. Fine vessel screening looks into local structural inconsistencies in vessels properties, into noise, or into not expected intensity variations observed inside pre-known vessel-body areas. The vessels are first modelled sufficiently but not precisely by their walls with a tubular model-structure that is the result of an initial segmentation. This provides a chart of likely Vessel Wall Pixels (VWPs) yielding a form of a likelihood vessel map mainly based on gradient filter's intensity and spatial arrangement parameters (e.g., linear consistency). Specific vessel parameters (centerline, width, location, fall-away rate, main orientation) are post-computed by convolving the image with a set of pre-tuned spatial filters called Matched Filters (MFs). These are easily computed as Gaussian-like 2D forms that use a limited range sub-optimal parameters adjusted to the dominant vessel characteristics obtained by Spatial Grey Level Difference statistics limiting the range of search into vessel widths of 16, 32, and 64 pixels. Sparse pixels are effectively eliminated by applying a limited range Hough Transform (HT) or region growing. Major benefits are limiting the range of parameters, reducing the search-space for post-convolution to only masked regions, representing almost 2% of the 2D volume, good speed versus accuracy/time trade-off. Results show the potentials of our approach in terms of time for detection ROC analysis and accuracy of vessel pixel (VP) detection.

No MeSH data available.


Related in: MedlinePlus