Local-Sensitive Connectivity Filter (LS-CF): A Post-Processing Unsupervised Improvement of the Frangi, Hessian and Vesselness Filters for Multimodal Vessel Segmentation

Abstract

A retinal vessel analysis is a procedure that can be used as an assessment of risks to the eye. This work proposes an unsupervised multimodal approach that improves the response of the Frangi filter, enabling automatic vessel segmentation. We propose a filter that computes pixel-level vessel continuity while introducing a local tolerance heuristic to fill in vessel discontinuities produced by the Frangi response. This proposal, called the local-sensitive connectivity filter (LS-CF), is compared against a naive connectivity filter to the baseline thresholded Frangi filter response and to the naive connectivity filter response in combination with the morphological closing and to the current approaches in the literature. The proposal was able to achieve competitive results in a variety of multimodal datasets. It was robust enough to outperform all the state-of-the-art approaches in the literature for the OSIRIX angiographic dataset in terms of accuracy and 4 out of 5 works in the case of the IOSTAR dataset while also outperforming several works in the case of the DRIVE and STARE datasets and 6 out of 10 in the CHASE-DB dataset. For the CHASE-DB, it also outperformed all the state-of-the-art unsupervised methods.

Keywords: Frangi filter; unsupervised learning; vessel segmentation; vesselness.

Figure 1

Overall variations of the Frangi filter gray-level response. (a) Image 11L from the CHASE-DB dataset. (b) Thicker vessel caliber response, which suits to the reality of the vessel, but contains several artefacts in non-vessel areas. (c) Smoother, less-noisy overall response but erases parts of important vessel branches. (d) Very noisy response that highlights most parts of the vessels but is more responsive to the edges of the vessel branches in contrast to their interior. (e) Less-noisy version of -e- but still with the same vessel caliber issue. (f) Less noisy version of -b- but looses more vessel information.

Figure 2

Image 02 from the test folder of the DRIVE dataset (a) and its ground truth (b). (a) Input image. (b) Ground truth.

Figure 3

The Frangi filter response (a) and its thresholded version at intensity 100 (b). (a) Frangi filter response. (b) Frangi filter response after threshold.

Figure 4

The connectivity filter output and its thresholded version. (a) Connectivity filter. (b) Thresholded version ($t>0$) of (a).

Figure 5

Improvements obtained with the locally sensitive version of the connectivity filter (LS-CF). (a) Image 01 of the DRIVE dataset. (b) Image 14 of the DRIVE dataset.

Figure 6

Steps and parameter analysis of the LS-CF algorithm. Image 14L of the CHASE-DB dataset. (a) The Frangi filter response. (b) The thresholded Frangi filter response shown in (a). (c) The LS-CF algorithm ($MA{X}_{score}=350$ and $MA{X}_{dist}=4$). (d) The thresholded version of the LS-CF response shown (c). (e) The LS-CF algorithm ($MA{X}_{score}=350$ and $MA{X}_{dist}=8$). (f) The thresholded version of the LS-CF response shown (e).

Figure 6

Steps and parameter analysis of the LS-CF algorithm. Image 14L of the CHASE-DB dataset. (a) The Frangi filter response. (b) The thresholded Frangi filter response shown in (a). (c) The LS-CF algorithm ($MA{X}_{score}=350$ and $MA{X}_{dist}=4$). (d) The thresholded version of the LS-CF response shown (c). (e) The LS-CF algorithm ($MA{X}_{score}=350$ and $MA{X}_{dist}=8$). (f) The thresholded version of the LS-CF response shown (e).

Figure 7

Results using an image from the DRIVE dataset. (a) Image 05—DRIVE dataset. (b) Ground truth. (c) Thresholded Frangi. (d) Gray-level CF response (proposal). (e) Thresholded CF response (proposal). (f) LS-CF response (proposal).

Figure 7

Results using an image from the DRIVE dataset. (a) Image 05—DRIVE dataset. (b) Ground truth. (c) Thresholded Frangi. (d) Gray-level CF response (proposal). (e) Thresholded CF response (proposal). (f) LS-CF response (proposal).

Figure 8

The worst numerical and visual segmentation result for the DRIVE dataset. (a) Image 08 from the DRIVE dataset. (b) Ground truth of image 08. (c) Thresholded Frangi response of image 08. (d) Gray-level CF response of image 08 (proposal). (e) Thresholded CF response of image 08 (proposal). (f) LS-CF response of image 08 (proposal).

Figure 8

The worst numerical and visual segmentation result for the DRIVE dataset. (a) Image 08 from the DRIVE dataset. (b) Ground truth of image 08. (c) Thresholded Frangi response of image 08. (d) Gray-level CF response of image 08 (proposal). (e) Thresholded CF response of image 08 (proposal). (f) LS-CF response of image 08 (proposal).

Figure 9

Visual performance of the LS-CF in a variety of datasets. (a) Image 02 from the IOSTAR dataset (retinal scanner laser ophthalmoscope). (b) Segmentation result of image 02 from the IOSTAR dataset using the proposed LS-CF. (c) Image 11R from the CHASE-DB dataset. (d) Segmentation result of image 11R from the CHASE-DB dataset using the proposed LS-CF. (e) Image 4 from the OSIRIX X-ray angiographic dataset. (f) Segmentation result of image 4 from the OSIRIX X-ray angiographic dataset using the proposed LS-CF.

Figure 9

Visual performance of the LS-CF in a variety of datasets. (a) Image 02 from the IOSTAR dataset (retinal scanner laser ophthalmoscope). (b) Segmentation result of image 02 from the IOSTAR dataset using the proposed LS-CF. (c) Image 11R from the CHASE-DB dataset. (d) Segmentation result of image 11R from the CHASE-DB dataset using the proposed LS-CF. (e) Image 4 from the OSIRIX X-ray angiographic dataset. (f) Segmentation result of image 4 from the OSIRIX X-ray angiographic dataset using the proposed LS-CF.