Fast automatic bone surface segmentation in ultrasound images without machine learning

Conference item

Reconstructing 3D bone images with 2D clinical ultrasound image is one of the primary developmental trends of computer-assisted orthopaedic surgery procedures, and real-time bone segmentation is required for such development. We previously presented a dynamic programming method with local phase tensor extraction for bone structure segmentation that could process one ultrasound frame with a true positive ratio of 71% in approximately 1 s. The present study aimed to reduce the segmentation time to enable real-time computational capacity for clinical application developments. A simplified bone probability algorithm was optimised by systematically identifying and removing the components which cost most computing resources. The segmentation results produced by the bone probability method were compared to the local phase method, and manual segmentation carried out by clinical experts. The proposed method had higher recall metric (0.67) than the local phase method (0.61), while the computational time is reduced to 0.02 s per image. However, the bone probability method did not perform as well as the local phase method in specificity and precision metrics. In conclusion, the simplified version of the segmentation algorithm improved computational speed and promised an advantage in further real time application developments, but additional functions that can improve accuracy and further extensive validations are still required before further clinical application developments.

Authors: Tu, SJ; Morel, J; Chen, M; Mellon, SJ

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: Springer
• Host title: Medical Image Understanding and Analysis (MIUA 2021)
• Pages: 250-264
• Series: Lecture Notes in Computer Science
• Series number: 12722
• Publication date: 2021-07-06
• Acceptance date: 2021-05-04
• Event title: 25th UK Conference on Medical Image Understanding and Analysis (MIUA 2021)
• Event location: Oxford, UK
• Event website: https://miua2021.com/
• Event start date: 2021-07-12
• Event end date: 2021-07-14
• DOI: 10.1007/978-3-030-80432-9_20
• EISSN: 1611-3349
• ISSN: 0302-9743
• EISBN: 9783030804329
• ISBN: 9783030804312
• Language: English
• Keywords: Networking and Information Technology R&D (NITRD); biomedical imaging; bioengineering; 46 information and computing sciences; machine learning and artificial intelligence