CN-122023446-A - Medical image focus segmentation method and system for radiology department

CN122023446ACN 122023446 ACN122023446 ACN 122023446ACN-122023446-A

Abstract

The invention discloses a medical image focus segmentation method and system for radiology department, and relates to the field of medical images. By constructing multi-window level inputs of brain window, subdural window and bone window, a bone mask and bone boundary probability map is generated by utilizing a bone tissue segmentation network, and feature suppression processing and boundary enhancement processing are respectively introduced in an encoder stage and a decoding stage so as to weaken the influence of high-density artifacts of a bone region and strengthen boundary expression of a bone-attached region. And further performing probability correction based on a bone mask on the initial bleeding probability map, reducing false detection of a bone region, and obtaining an accurate cerebral bleeding segmentation result. The system comprises an image access module, a multi-window level construction module, a bone priori generation module, a cerebral hemorrhage segmentation module, a probability correction module and a clinical quantification module, and can output structural indexes such as hemorrhage volume, centroid, diffusion direction and the like. The invention solves the problems of bone sticking false detection, edge blurring, small-volume focus missing detection and the like, and remarkably improves the accuracy of automatic segmentation of cerebral hemorrhage.

Inventors

YE WEN
ZHANG JIULONG
ZHAN YI

Assignees

上海市公共卫生临床中心

Dates

Publication Date: 20260512
Application Date: 20260206
Priority Date: 20251126

Claims (10)

1. A medical image lesion segmentation method for radiology department, comprising the steps of: acquiring an original DICOM image of craniocerebral CT, and generating brain window, subdural window and bone window images based on a preset window width window level to form a multichannel input image; the bone priori generation, the multi-channel input image is input into a bone tissue segmentation network, and a bone mask representing the position of a skull region and a bone boundary probability map representing the position of the inner side boundary of the skull are generated; the feature suppression processing is carried out, the bone mask is input into a feature suppression unit after multi-scale convolution processing, the feature suppression unit generates suppression weights based on the spatial distribution of the bone mask, and suppresses feature responses of corresponding bone regions in intermediate features of the cerebral hemorrhage segmentation network encoder; The boundary enhancement processing is carried out, in the decoding process of the cerebral hemorrhage segmentation network, the bone boundary probability map is input into a boundary enhancement unit, and the boundary enhancement unit carries out boundary enhancement convolution or deformable convolution on decoding characteristics close to a skull region according to the bone boundary probability map, so that the bleeding edge detection capability of a bone adjacent region is improved; the bleeding probability correction is carried out, suppression correction is carried out on probability values corresponding to bone areas in the initial bleeding probability map based on a bone mask, and false positives caused by high bone density are reduced; And generating a segmentation mask, namely generating a cerebral hemorrhage segmentation mask according to the corrected hemorrhage probability map.
2. A medical image lesion segmentation method according to claim 1, wherein the bone tissue segmentation network comprises an encoder, a decoder, and an edge-enhanced convolution module for enhancing skull edge features to support the generation of bone masks and bone boundary probability maps.
3. The medical image lesion segmentation method according to claim 1, wherein the feature suppression unit generates the suppression weight G based on the bone mask by: ; ; Wherein the method comprises the steps of Features of the bone mask obtained by multi-scale convolution, For the cerebral hemorrhage profile output by the encoder, Is a characteristic diagram after suppression.
4. The medical image focus segmentation method for radiology department according to claim 1, wherein the boundary enhancement unit comprises a deformable convolution module or a boundary attention module generated based on a bone boundary probability map, the deformable convolution module adapts a skull curved surface structure through learning offset to improve the spatial alignment capability of features close to a bone region, and the boundary attention module generates boundary response weights based on the bone boundary probability map to enhance the feature expression capability of a bleeding region and a bone boundary transition region.
5. The medical image focus segmentation method for radiology department according to claim 1, wherein the hemorrhage probability correction comprises the steps of multiplying a bone mask and an initial hemorrhage probability map according to positions to obtain correction terms, restraining the hemorrhage probability corresponding to a bone region based on the correction terms, reducing false detection caused by bone artifacts, and carrying out boundary smoothing on the corrected probability map according to neighborhood consistency to reduce isolated artifacts of adjacent bone regions.
6. A medical image lesion segmentation method according to claim 1, wherein the total loss function of the cerebral hemorrhage segmentation network comprises: ; Wherein the method comprises the steps of In the form of a bone mask, In the form of a map of the probability of bleeding, Is a weighting coefficient. .
7. The medical image focus segmentation method for radiology department according to claim 1, wherein the multi-channel input images are fused through a channel attention module, wherein the channel attention module generates channel weights according to response differences of brain windows, subdural windows and bone windows to promote feature complementation capability among different CT window levels, and the joint expression effect of local structures is enhanced based on a cross-channel interaction mechanism.
8. The method of claim 1, wherein the brain hemorrhage volume, centroid coordinates, diffusion directivity index and symmetry deviation index are calculated according to the brain hemorrhage segmentation mask, and the structured auxiliary diagnosis information including focus range, spatial position and quantization index is generated based on the above parameters.
9. A medical image lesion segmentation system for radiology department, the system for implementing a medical image lesion segmentation method for radiology department as set forth in any one of claims 1-8, comprising: The image access module is used for analyzing the DICOM image sequence and completing pixel reconstruction and normalization processing; the multi-window level construction module is used for generating brain window, subdural window and bone window images to form a multi-channel input image; The bone prior generation module is used for generating a bone mask and a bone boundary probability map through a bone tissue segmentation network; The cerebral hemorrhage segmentation module is used for executing coding feature extraction, feature suppression processing, boundary enhancement processing and decoding reasoning to generate an initial hemorrhage probability map; the probability correction module is used for correcting the initial bleeding probability map based on the bone mask so as to inhibit false detection of the bone region; The post-processing module is used for communicating domain analysis, artifact filtering and small focus enhancement; and the clinical parameter calculation module is used for calculating cerebral hemorrhage volume, centroid and diffusion index according to the final segmentation mask and generating auxiliary diagnosis information.
10. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the method steps of any of claims 1 to 9.

Description

Medical image focus segmentation method and system for radiology department Technical Field The invention relates to the field of medical images, in particular to a medical image focus segmentation method and system for radiology department. Background The traditional deep learning method based on the U-Net structure usually adopts a single window input and single semantic segmentation strategy, lacks explicit modeling and priori constraint on the skull structure, leads to difficult accurate distinction of the skull and bone-attached bleeding area of the model, has insufficient characteristic expression capability on the arc-shaped boundary and the high-density area of the bone, and further cannot effectively inhibit bone artifacts and ensure complete edge recognition of the bleeding area. Therefore, a medical image lesion segmentation method and system for radiology department is provided. Disclosure of Invention The invention mainly aims to provide a medical image focus segmentation method and a medical image focus segmentation system XX for radiology department, which can effectively solve the problems in the background technology. In order to achieve the above purpose, the invention adopts the technical proposal that, A medical image lesion segmentation method and system for radiology department, comprising the steps of: acquiring an original DICOM image of craniocerebral CT, and generating brain window, subdural window and bone window images based on a preset window width window level to form a multichannel input image; the bone priori generation, the multi-channel input image is input into a bone tissue segmentation network, and a bone mask representing the position of a skull region and a bone boundary probability map representing the position of the inner side boundary of the skull are generated; the feature suppression processing is carried out, the bone mask is input into a feature suppression unit after multi-scale convolution processing, the feature suppression unit generates suppression weights based on the spatial distribution of the bone mask, and suppresses feature responses of corresponding bone regions in intermediate features of the cerebral hemorrhage segmentation network encoder; The boundary enhancement processing is carried out, in the decoding process of the cerebral hemorrhage segmentation network, the bone boundary probability map is input into a boundary enhancement unit, and the boundary enhancement unit carries out boundary enhancement convolution or deformable convolution on decoding characteristics close to a skull region according to the bone boundary probability map, so that the bleeding edge detection capability of a bone adjacent region is improved; the bleeding probability correction is carried out, suppression correction is carried out on probability values corresponding to bone areas in the initial bleeding probability map based on a bone mask, and false positives caused by high bone density are reduced; And generating a segmentation mask, namely generating a cerebral hemorrhage segmentation mask according to the corrected hemorrhage probability map. Further, the bone tissue segmentation network includes an encoder, a decoder, and an edge-enhanced convolution module for enhancing the skull edge features to support the generation of bone masks and bone boundary probability maps. Further, the feature suppression unit generates a suppression weight G based on the bone mask, and the calculation method is as follows: ;; Wherein the method comprises the steps of Features of the bone mask obtained by multi-scale convolution,For the cerebral hemorrhage profile output by the encoder,Is a characteristic diagram after suppression. The boundary enhancement unit comprises a deformable convolution module or a boundary attention module generated based on a bone boundary probability map, wherein the deformable convolution module adapts to a skull curved surface structure through learning offset to improve the spatial alignment capability of features close to a bone region, and the boundary attention module generates boundary response weights based on the bone boundary probability map to enhance the feature expression capability of a bleeding region and a bone boundary transition region. The method comprises the steps of obtaining a correction term by multiplying a bone mask and an initial bleeding probability map according to positions, restraining bleeding probability corresponding to a bone region based on the correction term, reducing false detection caused by bone artifacts, and carrying out boundary smoothing on the corrected probability map according to neighborhood consistency, so that isolated artifacts of adjacent regions of the bone are reduced. Further, the total loss function of the cerebral hemorrhage segmentation network comprises: ; Wherein the method comprises the steps of In the form of a bone mask,In the form of a map of the probability of bleeding,Is a weighting coef