Deep Feature Masking and Homograph for Cranial and Larynx Ultrasound Image Stitching

OBJECTIVE: The narrow acoustic window of the skull, along with the occlusion of the thyroid cartilage and airway in the larynx, leads to the issue of limited field of view during the ultrasound imaging process. To address the issues of unstable feature extraction, difficult correspondence matching, and irregular boundaries in the image stitching process, this paper proposes a cranial and larynx ultrasound image stitching method based on the Deep Feature Masking Homography Network (DFMH-Net) framework. METHODS: Firstly, FasterNet is utilized to extract multi-scale deep features from unlabeled cranial-larynx ultrasound images. The correspondence between images is established through feature matching, and MaskNet is employed to generate deep feature masks. Subsequently, a homography matrix estimation network is constructed based on the masked weighted features to achieve image deformation. Then, the deformed images are fed into the designed stitching network, and the spatial and mask information is extracted and fused through multi-level convolutions to complete the image stitching. Finally, a content-aware rectangling network is introduced to eliminate distortions and output wide-field ultrasound images with regularization. RESULTS: Experimental results demonstrate that the proposed method outperforms other comparison methods in both visual results and evaluation metrics. CONCLUSION: The proposed method effectively overcomes the imaging limitations caused by occlusion of the skull and thyroid cartilage, and can fully present key structures.