Improving dense conditional random field for retinal vessel segmentation by discriminative feature learning and thin-vessel enhancement

As retinal vessels in color fundus images are thin and elongated structures, standard pairwise based random fields, which always suffer the "shrinking bias" problem, are not competent for such segmentation task. Recently, a dense conditional random field (CRF) model has been successfully used in retinal vessel segmentation. Its corresponding energy function is formulated as a linear combination of several unary features and a pairwise term. However, the hand-crafted unary features can be suboptimal in terms of linear models. Here we propose to learn discriminative unary features and enhance thin vessels for pairwise potentials to further improve the segmentation performance.

Our proposed method comprises four main steps: firstly, image preprocessing is applied to eliminate the strong edges around the field of view (FOV) and normalize the luminosity and contrast inside FOV; secondly, a convolutional neural network (CNN) is properly trained to generate discriminative features for linear models; thirdly, a combo of filters are applied to enhance thin vessels, reducing the intensity difference between thin and wide vessels; fourthly, by taking the discriminative features for unary potentials and the thin-vessel enhanced image for pairwise potentials, we adopt the dense CRF model to achieve the final retinal vessel segmentation. The segmentation performance is evaluated on four public datasets (i.e. DRIVE, STARE, CHASEDB1 and HRF).

Experimental results show that our proposed method improves the performance of the dense CRF model and outperforms other methods when evaluated in terms of F1-score, Matthews correlation coefficient (MCC) and G-mean, three effective metrics for the evaluation of imbalanced binary classification. Specifically, the F1-score, MCC and G-mean are 0.7942, 0.7656, 0.8835 for the DRIVE dataset respectively; 0.8017, 0.7830, 0.8859 for STARE respectively; 0.7644, 0.7398, 0.8579 for CHASEDB1 respectively; and 0.7627, 0.7402, 0.8812 for HRF respectively.

The discriminative features learned in CNNs are more effective than hand-crafted ones. Our proposed method performs well in retinal vessel segmentation. The architecture of our method is trainable and can be integrated into computer-aided diagnostic (CAD) systems in the future.