CN-121982042-A - Mandibular duct segmentation method, mandibular duct segmentation system, electronic equipment and storage medium

CN121982042ACN 121982042 ACN121982042 ACN 121982042ACN-121982042-A

Abstract

The application discloses a mandibular duct segmentation method, a mandibular duct segmentation system, electronic equipment and a storage medium, wherein the mandibular duct segmentation method comprises the steps of acquiring an oral and maxillofacial image; cutting an oromandibular facial image to obtain a plurality of local images of a mandible, inputting the plurality of local images into a mandibular segmentation model, outputting the segmented images to obtain a segmentation result, wherein the mandibular segmentation model comprises an encoder, a decoder and a feature fusion module, three two-dimensional convolution branches are configured on three orthogonal projection planes in the feature fusion module, each two-dimensional convolution branch can acquire boundary information, central trend and local bending features of the mandible on different orthogonal planes, and fuse the boundary information, central trend and local bending features to obtain multi-view features, fuse the multi-view features and the local texture features to obtain enhancement features, and the decoder fuses the local texture features and the enhancement features and outputs the segmentation result. Through the feature fusion module, the whole coverage of the multi-view geometric features is realized, and the problems of segmentation fracture and boundary blurring caused by single feature extraction are effectively avoided.

Inventors

WANG SUYU
CUI YUQING
LI XIAOGUANG
ZHANG YU
ZHU NING
LU JUNXI
LIANG SHUYING

Assignees

北京工业大学
北京大学口腔医学院

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. A mandibular duct segmentation method, comprising: acquiring an oromaxillofacial image comprising mandibular and surrounding jaw anatomy; Cutting the oral and maxillofacial image, and screening to obtain a plurality of local images of mandibles; Inputting the plurality of partial images into a mandibular duct segmentation model, and outputting to obtain a segmentation result comprising a mandibular duct feature map; The mandibular duct segmentation model comprises an encoder, a decoder and a bottleneck layer connected between the encoder and the decoder, wherein a feature fusion module is arranged in the bottleneck layer, three two-dimensional convolution branches are configured on three orthogonal projection planes in the feature fusion module, each two-dimensional convolution branch can acquire boundary information, central trend and local bending features of a mandibular duct on different orthogonal planes, the feature fusion module fuses the boundary information, the central trend and the local bending features to obtain multi-view features, fuses the multi-view features and local texture features output by the encoder to obtain enhancement features, and the decoder fuses the local texture features output by the encoder and the enhancement features to output a mandibular duct feature map.
2. The mandibular duct segmentation method according to claim 1, wherein a mann bar module is further arranged in the bottleneck layer, the mann bar module carries out sequence modeling on the enhanced features output by the feature fusion module, acquires the continuous association features of the mandibular duct structure, and outputs the fusion features of the enhanced features and the continuous association features of the mandibular duct structure; The sequence modeling of the enhanced features output by the feature fusion module comprises the following steps: performing dimension remolding on the feature map corresponding to the enhanced feature to form a feature sequence; Performing linear projection and gating screening on the characteristic sequences to obtain effective characteristic sequences; Based on the effective feature sequence, acquiring continuous association features of the mandibular duct structure; And fusing the continuous association characteristic of the mandibular duct structure with the enhancement characteristic to obtain the fusion characteristic.
3. The mandibular duct segmentation method according to claim 2, wherein a hybrid multi-scale residual module is provided within the decoder, the hybrid multi-scale residual module comprising 7 x 7 convolution branches, 3 x 3 convolution branches, and 1 x1 convolution branches; the method comprises the steps of processing the 7X 7 convolution branch to obtain a global spatial structure association feature of a mandibular pipe, respectively inputting the global spatial structure association feature of the mandibular pipe into the 3X 3 convolution branch and the 1X 1 convolution branch to correspondingly obtain a local feature and a semantic feature, adding and fusing the global spatial structure association feature of the mandibular pipe, the local feature and the semantic feature element by element, and then superposing the global spatial structure association feature of the mandibular pipe with the local texture feature through residual connection, and outputting to obtain a mandibular pipe feature map.
4. The mandibular duct segmentation method according to claim 1, wherein the acquiring of the oromaxillofacial image including mandibular duct and surrounding jaw anatomy comprises: acquiring cone beam computed tomography image data comprising mandibular and surrounding jaw anatomy; And labeling all the image data, and labeling the actual position, boundary and form of the mandibular duct to obtain the oral and maxillofacial image.
5. The mandibular duct segmentation method according to claim 1, wherein the cropping the oromaxillofacial image, and the filtering to obtain a plurality of partial images of the mandible comprises: cutting the oral and maxillofacial image, and screening to obtain a plurality of initial partial images of the mandible area; resampling the plurality of initial partial images to obtain a plurality of initial partial images with preset sizes; And carrying out gray scale normalization on the plurality of initial partial images to obtain a plurality of partial images.
6. The mandibular duct segmentation method according to claim 1, wherein the encoder performs a multi-layer convolution and downsampling operation on the plurality of local images to obtain the local texture features.
7. The mandibular duct segmentation method according to claim 1, wherein three two-dimensional convolution branches are configured on three orthogonal projection planes in the feature fusion module, and the three orthogonal projection planes are a wide-high plane, a high-deep plane, and a wide-deep plane of a local texture feature map corresponding to local texture features output by the encoder.
8. A mandibular duct segmentation system, comprising: the acquisition module is used for acquiring an oral and maxillofacial image comprising a mandibular duct and surrounding jaw anatomy structures; The processing module is used for cutting the oral and maxillofacial images and screening to obtain a plurality of local images of the mandible; the execution module is used for inputting the plurality of partial images into a mandibular duct segmentation model and outputting a segmentation result comprising a mandibular duct feature map; The mandibular duct segmentation model comprises an encoder, a decoder and a bottleneck layer connected between the encoder and the decoder, wherein a feature fusion module is arranged in the bottleneck layer, three two-dimensional convolution branches are configured on three orthogonal projection planes in the feature fusion module, each two-dimensional convolution branch can acquire boundary information, central trend and local bending features of a mandibular duct on different orthogonal planes, the feature fusion module fuses the boundary information, the central trend and the local bending features to obtain multi-view features, fuses the multi-view features and local texture features output by the encoder to obtain enhancement features, and the decoder fuses the local texture features output by the encoder and the enhancement features to output a mandibular duct feature map.
9. An electronic device comprising at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the method of any of claims 1-7.
10. A storage medium storing a computer program executable by an electronic device, which when run on the electronic device causes the electronic device to perform the method of any one of claims 1-7.

Description

Mandibular duct segmentation method, mandibular duct segmentation system, electronic equipment and storage medium Technical Field The application relates to the technical field of artificial intelligence model cooperation, in particular to a mandibular duct segmentation method, a mandibular duct segmentation system, electronic equipment and a storage medium. Background Cone beam computed tomography (CBCT, cone Beam Computed Tomography) has been widely used for diagnosis and treatment planning of oromaxillofacial structures, by virtue of its significant advantages of low cost, low radiation dose, and high resolution. The mandible is used as a key anatomical structure in the jawbone to accommodate important tissues such as lower alveolar nerves and blood vessels, and the space position and the running form of the mandible are directly related to the safety and success rate of operations such as oral implantation, resist tooth extraction and the like, so that the mandible is accurately and stably automatically segmented in a pre-operation CBCT image, and the mandible has important significance in reducing operation risk and improving clinical decision efficiency. Currently, convolutional neural networks, transformers, and other architectures have been applied to mandibular segmentation tasks. However, the existing architecture for mandibular duct segmentation task is difficult to adaptively capture morphological changes in different directions and scales, the segmentation result is easy to generate topology fracture, omission or boundary overflow, the requirement of high-precision clinical operation planning is difficult to be met, and the segmentation stability is insufficient in the area with severe running change or low contrast. Disclosure of Invention In order to overcome the defects in the prior art, the application provides a mandibular duct segmentation method, a mandibular duct segmentation system, electronic equipment and a storage medium. In order to achieve the above object, the present application provides a mandibular duct segmentation method, comprising: acquiring an oromaxillofacial image comprising mandibular and surrounding jaw anatomy; Cutting out the oral and maxillofacial images, and screening to obtain a plurality of local images of mandibles; inputting a plurality of partial images into a mandibular duct segmentation model, and outputting to obtain a segmentation result comprising a mandibular duct feature map; The mandibular duct segmentation model comprises an encoder, a decoder and a bottleneck layer connected between the encoder and the decoder, wherein a feature fusion module is arranged in the bottleneck layer, three two-dimensional convolution branches are configured on three orthogonal projection planes in the feature fusion module, each two-dimensional convolution branch can acquire boundary information, center trend and local bending features of the mandibular duct on different orthogonal planes, the feature fusion module fuses the boundary information, the center trend and the local bending features to obtain multi-view features, and fuses the multi-view features and the local texture features output by the encoder to obtain enhancement features, and the decoder fuses the local texture features and the enhancement features output by the encoder to obtain a mandibular duct feature map. Further, a Manba module is further arranged in the bottleneck layer, the Manba module carries out sequence modeling on the enhanced features output by the feature fusion module, continuous association features of the mandibular duct structure are obtained, and fusion features of the enhanced features and the continuous association features of the mandibular duct structure are output; the sequence modeling of the enhanced features output by the feature fusion module comprises the following steps: Performing dimension remolding on the feature map corresponding to the enhanced feature to form a feature sequence; performing linear projection and gating screening on the feature sequences to obtain effective feature sequences; Based on the effective feature sequence, acquiring continuous association features of the mandibular duct structure; and fusing the continuous association characteristic of the mandibular duct structure with the enhancement characteristic to obtain a fusion characteristic. Further, a mixed multi-scale residual error module is arranged in the decoder, hybrid multi-scale residual error module comprising 7X 7 convolution branches comprising 7X 7 convolution branch; The method comprises the steps of processing enhancement features by 7X 7 convolution branches to obtain global spatial structure related features of a mandibular pipe, respectively inputting the global spatial structure related features of the mandibular pipe into 3X 3 convolution branches and 1X 1 convolution branches to correspondingly obtain local features and semantic features, adding and fusing the global spatial st