A discriminatively trained fully connected conditional random field model for blood vessel segmentation in fundus images

JOSÉ IGNACIO ORLANDO; ELENA PROKOFYEVA; MATTHEW B. BLASCHKO

IEEE TRANSACTIONS ON BIO-MEDICAL ENGINEERING

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC

Lugar: New York; Año: 2017 vol. 64 p. 16 - 27

0018-9294

Objective: In this work, we present an extensivedescription and evaluation of our method for blood vesselsegmentation in fundus images based on a discriminativelytrained, fully connected conditional random field model. Methods:Standard segmentation priors such as a Potts model or totalvariation usually fail when dealing with thin and elongatedstructures. We overcome this difficulty by using a conditionalrandom field model with more expressive potentials, takingadvantage of recent results enabling inference of fully connectedmodels almost in real-time. Parameters of the method are learnedautomatically using a structured output support vector machine,a supervised technique widely used for structured predictionin a number of machine learning applications. Results: Ourmethod, trained with state of the art features, is evaluatedboth quantitatively and qualitatively on four publicly availabledata sets: DRIVE, STARE, CHASEDB1 and HRF. Additionally,a quantitative comparison with respect to other strategies isincluded. Conclusion: The experimental results show that thisapproach outperforms other techniques when evaluated in termsof sensitivity, F1-score, G-mean and Matthews correlation coefficient. Additionally, it was observed that the fully connectedmodel is able to better distinguish the desired structures than thelocal neighborhood based approach. Significance: Results suggestthat this method is suitable for the task of segmenting elongatedstructures, a feature that can be exploited to contribute withother medical and biological applications.