Deep learning for coronary artery segmentation in X-ray angiograms using a patch-based training
Abstract — This paper presents a new method for coronary artery segmentation in X-ray angiograms based on deep learning and a patch-based training.
The blood vessel segmentation is performed using the U-Net convolutional neural network, which has been trained using patches extracted from the original
angiograms instead of using complete images. The publicly available database of coronary angiograms DCA1 containing 130 angiograms with their respective ground-truth has been
used to generate the training patterns and subsequently to evaluate and compare the segmentation performance of the proposed method.
The hyper-parameter configuration used for training the U-Net parameters has been selected from 90 possible combinations according to five binary classification metrics.
Each combination involving the selection of a patch size, weight assigned to the blood vessel class, and learning rate used by the optimization method, has been used in order
to train the U-Net parameters with patterns extracted from a set of 100 images. The segmentation performance of the proposed method is compared with five specialized blood vessel
segmentation methods from the state of the art using a test set of 30 images, achieving the highest accuracy (0.977) and Dice similarity coefficient (0.779). Moreover, the experimental
results have also shown that the proposed method is suitable to be integrated into a computer-aided system to support decision making in medical practice.
Authors:
Fernando Cervantes-Sanchez, Ivan Cruz-Aceves, Arturo Hernandez-Aguirre,Martha A. Hernandez-Gonzalez and Sergio E. Solorio-Meza
Research paper submitted to 16th International Symposium on Medical Information Processing and Analysis (SIPAIM-2020- PERU)
Experimental results
Training and validation performance, in terms of the weighted cross entropy through epochs. Only the top three U-Net hyper-parameter combinations, and the best configuration that use complete images, are presented.
Binary classification accuracy and Dice similarity coefficient for the blood vessel segmentation performance, evaluated using the test set of 30 angiograms.
Comparative analysis of the segmentation performance of 5 state-of-the-art blood vessel segmentation methods, and the proposed method using the test of 30 angiograms.
