Accelerating Convolutional Sparse Coding for Curvilinear Structures Segmentation by Refining SCIRD-TS Filter Banks

• Published in: IEEE transactions on medical imaging Vol. 35; no. 11; pp. 2381 - 2392
• Main Authors: Annunziata, Roberto, Trucco, Emanuele
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.11.2016; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Acceleration; Algorithms; Axons; Biomedical imaging; Blood vessels; Coding; Convolutional codes; Convolutional sparse coding; Detectors; Encoding; Filter banks; Filters; Humans; Image Interpretation, Computer-Assisted - methods; Image Processing, Computer-Assisted - methods; Image segmentation; Machine learning; Neural Networks (Computer); neurites; Neurites - physiology; Retina; retinal blood vessels; Retinal Vessels - diagnostic imaging; segmentation; Spherical coordinates
• ISSN: 0278-0062; 1558-254X; 1558-254X
• DOI: 10.1109/TMI.2016.2570123

Deep learning has shown great potential for curvilinear structure (e.g., retinal blood vessels and neurites) segmentation as demonstrated by a recent auto-context regression architecture based on filter banks learned by convolutional sparse coding. However, learning such filter banks is very time-consuming, thus limiting the amount of filters employed and the adaptation to other data sets (i.e., slow re-training). We address this limitation by proposing a novel acceleration strategy to speed-up convolutional sparse coding filter learning for curvilinear structure segmentation. Our approach is based on a novel initialisation strategy (warm start), and therefore it is different from recent methods improving the optimisation itself. Our warm-start strategy is based on carefully designed hand-crafted filters (SCIRD-TS), modelling appearance properties of curvilinear structures which are then refined by convolutional sparse coding. Experiments on four diverse data sets, including retinal blood vessels and neurites, suggest that the proposed method reduces significantly the time taken to learn convolutional filter banks (i.e., up to -82%) compared to conventional initialisation strategies. Remarkably, this speed-up does not worsen performance; in fact, filters learned with the proposed strategy often achieve a much lower reconstruction error and match or exceed the segmentation performance of random and DCT-based initialisation, when used as input to a random forest classifier.