|
Optic disk detection and segmentation approaches based on vessel network |
|---|---|
| รหัสดีโอไอ | |
| Title | Optic disk detection and segmentation approaches based on vessel network |
| Creator | Nittaya Muangnak |
| Contributor | Pakinee Aimmanee, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2559 |
| Keyword | Optic disk detection, Optic disk localization, Optic disk segmentation, Vessel network, Vessel transform, Vessel vector based phase portrait analysis, Hybrid approach |
| Abstract | Precise localization of the optic disk (OD) in retinal images is one of the challenges in ophthalmic image processing. Although many efforts have been made towards finding automated numerical solutions, they often fail on retinal images characterized by poor quality. Therefore, we introduce three novel methods namely, Vessel Transform (VT), Vessel Vector based Phase Portrait Analysis (VVPPA), and Hybrid Approach (HA) for automatic detection of the OD based primarily on retinal blood vessels. To localize OD, VT finds a set of solution in an image space that yields the smallest sum of distance from a solution point to clusters of vessels. VVPPA uses convergence points obtained from Phase Portrait Analysis (PPA) operated on vectors derived from vessels to get the location of OD. The HA uses a set of rules obtained from a decision tree to alternate using VT and VVPPA. These three methods are integrated with the scale space approach (SS) to obtain the OD boundary. The integration of VT, VVPPA, and HA with SS is defined by SSVT, SSVVPPA, and SSHA, respectively. The new algorithms have been tested against the existing methods: fuzzy convergence and circular transform. The numerical experiments demonstrate that our proposed algorithm outperforms the existing methods, especially on poor quality images. For localization of the OD, the HA gets the highest accuracy of 98% regardless of poor image quality and satisfactorily comparable with the existing method for the fair ones. For OD segmentation, all approaches are generally better than the existing methods.Specifically, SSVT obtains the highest positive predictive value (PPV) of 79.22% while SSHA gets the highest sensitivity of 53.05%, respectively for poor quality images. |
