Medical image lesion segmentation method and system for radiology department

Abstract

The invention discloses a medical image lesion segmentation method and system for a radiology department, and relates to the field of medical images. A bone mask and a bone boundary probability graph are generated by constructing multi-window input of a brain window, a subdural window and a bone window and utilizing a bone tissue segmentation network, and feature suppression processing and boundary enhancement processing are introduced in an encoder stage and a decoding stage respectively so as to weaken the influence of high-density artifacts in a bone region and strengthen boundary expression of a bone attachment region. And further executing probability correction based on a bone mask on the initial hemorrhage probability graph, reducing bone region false detection, and obtaining an accurate cerebral hemorrhage segmentation result. The system comprises an image access module, a multi-window construction module, a bone prior generation module, a cerebral hemorrhage segmentation module, a probability correction module and a clinical quantification module, and can output structured indexes such as hemorrhage volume, mass center and diffusion direction. According to the method, the problems of bone sticking false detection, fuzzy edge, small-size focus missing detection and the like are solved, and the accuracy of automatic cerebral hemorrhage segmentation is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of medical imaging, and in particular to a method and system for lesion segmentation in medical images used in radiology. Background Technology

[0002] Cranial CT is the preferred imaging modality for the clinical diagnosis of cerebral hemorrhage. However, automatic segmentation of cerebral hemorrhage still faces key technical challenges, including high false positive rates in bone-adjacent areas, complex hemorrhage margin morphology, easy neglect of small lesions, underutilization of complementary information from multi-window images, and significant interference from bone artifacts. Traditional deep learning methods based on U-Net-like structures often employ single-window input and single semantic segmentation strategies, lacking explicit modeling and prior constraints on the skull structure. This results in the model's inability to accurately distinguish between the skull and hemorrhage areas adjacent to the bone, insufficient ability to represent features of curved bone boundaries and high-density regions, and consequently, inability to effectively suppress bone artifacts and ensure complete edge recognition of the hemorrhage area. Therefore, this paper proposes a medical image lesion segmentation method and system for radiology. Summary of the Invention

[0003] The main objective of this invention is to provide a method and system for lesion segmentation in medical images for radiology, which can effectively solve the problems in the background art.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A method and system for lesion segmentation in medical images used in radiology, comprising the following steps:

[0006] Multi-window image construction: acquire the original DICOM image of cranial CT, and generate brain window, subdural window and bone window images based on preset window width and window level to form a multi-channel input image;

[0007] Bone prior generation involves inputting the multi-channel input image into a bone tissue segmentation network to generate a bone mask representing the location of the skull region and a bone boundary probability map representing the location of the inner boundary of the skull.

[0008] Feature suppression processing involves inputting the bone mask 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 the feature responses of the corresponding bone regions in the intermediate features of the brain hemorrhage segmentation network encoder.

[0009] Boundary enhancement processing: During the decoding process of the brain hemorrhage segmentation network, the bone boundary probability map is input into the boundary enhancement unit. The boundary enhancement unit performs boundary enhancement convolution or deformable convolution on the decoding features of the region near the skull according to the bone boundary probability map, thereby improving the detection capability of hemorrhage edges in the bone-adjacent region.

[0010] Bleeding probability correction is performed based on bone mask to suppress and correct the probability values ​​corresponding to bone regions in the initial bleeding probability map, thereby reducing false positives caused by high bone density.

[0011] Segmentation mask generation: A brain hemorrhage segmentation mask is generated based on the corrected hemorrhage probability map.

[0012] Furthermore, the bone tissue segmentation network includes an encoder, a decoder, and an edge enhancement convolution module, which is used to enhance the edge features of the skull to support the generation of bone masks and bone boundary probability maps.

[0013] Furthermore, the feature inhibition unit generates an inhibition weight G based on the bone mask, which is calculated as follows:

[0014] G=sigmoid[Conv(BoneFeat)]; F′=F ICH ⊙(1-G);

[0015] Where BoneFeat represents the features obtained from the bone mask through multi-scale convolution, F ICh F is the brain hemorrhage feature map output by the encoder, and F is the suppressed feature map.

[0016] Furthermore, the boundary enhancement unit includes a deformable convolution module or a boundary attention module generated based on a bone boundary probability map; the deformable convolution module adapts to the curved surface structure of the skull by learning offsets, thereby improving the spatial alignment capability of features near the bone region; the boundary attention module generates boundary response weights based on the bone boundary probability map, thereby enhancing the feature representation capability of the transition region between the bleeding region and the bone boundary.

[0017] Furthermore, the bleeding probability correction includes: multiplying the bone mask by the initial bleeding probability map according to position to obtain a correction term; suppressing the bleeding probability corresponding to the bone region based on the correction term to reduce false detections caused by bone artifacts; and smoothing the boundary of the corrected probability map according to neighborhood consistency to reduce isolated artifacts in the bone's adjacent region.

[0018] Furthermore, the total loss function of the brain hemorrhage segmentation network includes:

[0019] L = L Dice +L CE +λ∑(BoneMask×P ICH );

[0020] BoneMask is the bone mask, P ICH This is a bleeding probability plot, where λ is the weighting coefficient.

[0021] Furthermore, the multi-channel input images are fused by a channel attention module; the channel attention module generates channel weights based on the response differences of brain windows, subdural windows and bone windows to enhance the feature complementarity between different CT window positions; and enhances the joint expression effect of local structures based on a cross-channel interaction mechanism.

[0022] Furthermore, the brain hemorrhage volume, centroid coordinates, diffusion direction index, and symmetry offset index are calculated based on the brain hemorrhage segmentation mask; and structured auxiliary diagnostic information containing lesion range, spatial location, and quantitative indicators is generated based on the above parameters.

[0023] A medical image lesion segmentation system for radiology, the system being used in a method for medical image lesion segmentation in radiology, comprising:

[0024] Image input module: used to parse DICOM image sequences and perform pixel reconstruction and normalization processing;

[0025] Multi-window construction module: used to generate brain window, subdural window and bone window images to form a multi-channel input image;

[0026] Bone Prior Generation Module: Used to generate bone masks and bone boundary probability maps through a bone tissue segmentation network;

[0027] The brain hemorrhage segmentation module is used to perform coded feature extraction, feature suppression processing, boundary enhancement processing, and decoding inference to generate an initial hemorrhage probability map.

[0028] Probability correction module: used to correct the initial hemorrhage probability map based on the bone mask in order to suppress false detections of bone regions;

[0029] Post-processing module: used for connected component analysis, artifact filtering, and enhancement of small lesions;

[0030] Clinical parameter calculation module: used to calculate the volume, centroid, and diffusion index of cerebral hemorrhage based on the final segmentation mask and generate auxiliary diagnostic information.

[0031] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described.

[0032] The present invention has the following beneficial effects:

[0033] Compared with existing technologies, this solution introduces bone prior information composed of bone masks and bone boundary probability maps, which are used for feature suppression, boundary enhancement, and probability correction in the segmentation process, respectively. This achieves explicit weakening of high-density responses in the skull region and structural enhancement of bleeding edges in adjacent bone regions. Combined with multi-channel input composed of multi-window images and a channel attention mechanism, the model can adaptively utilize complementary information between brain windows, subdural windows, and bone windows to improve the ability to distinguish tissues of different densities. The multi-scale bone structure suppression strategy, deformable convolutional structure guided by bone boundaries, and subsequent probability correction mechanism proposed in this invention can significantly reduce false detections caused by bone artifacts, improve the segmentation accuracy of regions with blurred boundaries and adherence to bone, and maintain stable performance in complex scenarios such as small-volume bleeding and arc-shaped boundary bleeding. Combined with the output structured clinical quantitative parameters, the model is significantly improved in terms of algorithm interpretability, clinical readability, and practical deployment reliability, thereby solving the performance bottleneck of traditional segmentation methods in scenarios with false detections of adherence to bone, high-density interference, and blurred edges. Attached Figure Description

[0034] Figure 1 This is an overall flowchart of the present invention;

[0035] Figure 2 This is a block diagram of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0037] Example

[0038] System Overall Process

[0039] The brain hemorrhage segmentation method in this embodiment includes seven stages: multi-window image construction, bone prior generation, feature suppression, boundary enhancement, probability correction, segmentation mask generation, and clinical quantitative output. The process is as follows: Figure 1 As shown.

[0040] Multi-window image construction

[0041] Acquire the raw DICOM cranial CT image sequence output from the emergency department or radiology department, and input it into the image access module to complete the following process:

[0042] DICOM parsing: Reads metadata such as pixel matrix, slice thickness, window width and window level from head CT scans;

[0043] Sequence reconstruction: Sort slices according to instance Slice Location to construct complete 3D volume data;

[0044] Normalization and preprocessing: The HU values ​​are scaled to the range of [-1, 1] using linear normalization;

[0045] Multi-window generation: Generates three types of window level images based on commonly used clinical settings:

[0046] Brain window (WW=80, WL=40); subdural window (WW=130, WL=80); bone window (WW=2800, WL=600);

[0047] The three types of window images are stitched together by channel to form a three-channel input image, which is used for subsequent network processing.

[0048] Bone prior generation

[0049] The constructed three-channel image is input into the bone tissue segmentation network. The network adopts a U-Net structure, including a 4-layer encoder, a 4-layer decoder, and a skip connection structure. An edge enhancement convolutional module is introduced at the end of the encoder to enhance the skull boundary.

[0050] Through network reasoning, we obtained the following:

[0051] BoneMask: Reflects the binarization probability of the skull region through Sigmoid output;

[0052] BoneBoundary: Enhances the output of continuous boundary probabilities through convolution with specific edges.

[0053] The bone boundary probability map primarily enhances the high-gradient region on the inner side of the skull for boundary enhancement in the subsequent decoding stage.

[0054] Feature suppression processing

[0055] To suppress high-density false detections caused by bone artifacts, the bone mask is subjected to multi-scale convolution (kernel sizes of 3, 5, and 7) to extract hierarchical bone features, resulting in a bone feature map (BoneFeat). This map is then fed into a feature suppression unit, and the suppression weight G is calculated.

[0056] G=sigmoid[Conv(BoneFeat)];

[0057] Where Conv represents a 1×1 convolutional layer, used to compress the channel dimension.

[0058] Then, based on the following formula, the intermediate features F of the encoder are... ICH Suppression treatment:

[0059] F′=F ICH ⊙(1-G);

[0060] Where F ′The image shows the suppressed feature map. The suppression operation significantly reduces the network's response to high-density bone structures in the bone region, preventing the model from misidentifying bone as bleeding.

[0061] Boundary enhancement processing

[0062] During the decoding phase, the bone boundary probability map is input into the boundary enhancement unit, and one of the following two enhancement methods is performed (which can be dynamically selected):

[0063] 1. Deformable convolutional enhancement

[0064] Migration field Δ generated based on bone boundary probability map P It is used to guide the convolution kernel to adapt to the curved surface structure of the skull, thereby enhancing the feature representation of the adjacent bone region.

[0065] 2. Boundary attention mechanism

[0066] An attention weight map A is generated based on the bone boundary probability, and the decoded features F are then processed. dec Enhancement:

[0067] F enh =F dec ⊙(1+A);

[0068] Enhanced features provide clearer detail of the boundary in the bone-adjacent region, improving the recall rate of suture bleeding.

[0069] Bleeding probability correction

[0070] After decoding, the initial bleeding probability map P is obtained. TCH Then, the probability of execution is corrected based on the bone mask:

[0071] 1. Bone region inhibition

[0072] P ′ =P ICH ×(1-BoneMask);

[0073] 2. Neighborhood smoothing

[0074] 3×3 median filtering or bilateral filtering is used to reduce isolated artifacts at bone boundaries and ensure spatial consistency of the probability map.

[0075] The corrected probability map significantly reduces false positives caused by bone artifacts and more closely approximates the actual hemorrhage distribution.

[0076] Segmentation mask generation

[0077] The corrected probability map P ′ Binarization is performed using a threshold of 0.5 to generate the final brain hemorrhage segmentation mask:

[0078] Canceling regions with a connected component area of ​​less than 10 pixels is performed to reduce noise artifacts;

[0079] 3D morphological closing operations are used to eliminate holes and improve segmentation coherence.

[0080] Clinical quantitative parameter output

[0081] Based on the generated final segmentation mask, the following quantization metrics are calculated:

[0082] Cerebral hemorrhage volume (m):

[0083] V = ∑ i (A i ×SliceThickness×PixelSpacing 2 );

[0084] Centroid coordinates: used to determine the location and degree of displacement of the lesion.

[0085] Diffusion directionality index: Based on principal component analysis (PCA) to statistically determine the direction of bleeding shape.

[0086] Symmetry offset index: assesses whether the lesion crosses the midline.

[0087] train

[0088] To achieve effective learning of the bleeding segmentation network, this embodiment adopts the following training process:

[0089] Training data:

[0090] Data from 1200 cranial CT scans were collected, including cases of acute hemorrhage, subacute hemorrhage, and some cases of chronic hemorrhage.

[0091] Labeling method:

[0092] The labeling was performed layer by layer by two radiologists with more than 8 years of experience.

[0093] Loss function:

[0094] The following total loss is adopted:

[0095] L = L Dice +L CE +λ∑(BoneMask×P ICH );

[0096] Among them, L Dice Enhance boundary consistency; L CE Improve overall classification performance; exclusion terms reduce false positives in bone artifacts; λ = 0.1; optimizer: Adam, initial learning rate 1e-4.

[0097] Training strategy: Random rotation ±15°; random translation 0–10 pixels; random artifact simulation to enhance model robustness.

[0098] The model converged after 200 epochs.

[0099] Inference Deployment

[0100] The system of this invention can be deployed on: hospital PACS servers; edge inference devices (NVIDIA Jetson series); and cloud-based deep inference platforms.

[0101] The reasoning process takes only 0.4–0.9 seconds per case, which can meet the needs of emergency department for immediate diagnosis.

[0102] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for lesion segmentation in medical images used in radiology, characterized in that, Includes the following steps: Multi-window image construction: acquire the original DICOM image of cranial CT, and generate brain window, subdural window and bone window images based on preset window width and window level to form a multi-channel input image; Bone prior generation involves inputting the multi-channel input image into a bone tissue segmentation network to generate a bone mask representing the location of the skull region and a bone boundary probability map representing the location of the inner boundary of the skull. Feature suppression processing involves inputting the bone mask 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 the feature responses of the corresponding bone regions in the intermediate features of the brain hemorrhage segmentation network encoder. Boundary enhancement processing: During the decoding process of the brain hemorrhage segmentation network, the bone boundary probability map is input into the boundary enhancement unit. The boundary enhancement unit performs boundary enhancement convolution or deformable convolution on the decoding features of the region near the skull according to the bone boundary probability map, thereby improving the detection capability of hemorrhage edges in the bone-adjacent region. Bleeding probability correction is performed based on bone mask to suppress and correct the probability values ​​corresponding to bone regions in the initial bleeding probability map, thereby reducing false positives caused by high bone density. Segmentation mask generation: A brain hemorrhage segmentation mask is generated based on the corrected hemorrhage probability map.

2. The method for lesion segmentation in medical images for radiology according to claim 1, characterized in that, The bone tissue segmentation network includes an encoder, a decoder, and an edge enhancement convolution module. The edge enhancement convolution module is used to enhance the edge features of the skull to support the generation of bone masks and bone boundary probability maps.

3. The method for lesion segmentation in medical images for radiology according to claim 1, characterized in that, The feature inhibition unit generates an inhibition weight G based on the bone mask, and its calculation method is as follows: G=sigmoid[Conv(BoneFeat)];F ′ =F ICH ⊙(1-G); Where BoneFeat represents the features obtained from the bone mask through multi-scale convolution, F ICH F' represents the brain hemorrhage feature map output by the encoder, and F' represents the suppressed feature map.

4. The method for lesion segmentation in medical images for radiology according to claim 1, characterized in that, The boundary enhancement unit includes a deformable convolution module or a boundary attention module generated based on a bone boundary probability map. The deformable convolution module adapts to the curved surface structure of the skull by learning offsets, thereby improving the spatial alignment capability of features near the bone region; the boundary attention module generates boundary response weights based on the bone boundary probability map, thereby enhancing the feature representation capability of the transition region between the bleeding region and the bone boundary.

5. A method for lesion segmentation in medical images for radiology according to claim 1, characterized in that, The bleeding probability correction includes: multiplying the bone mask by the initial bleeding probability map according to position to obtain a correction term; suppressing the bleeding probability corresponding to the bone region based on the correction term to reduce false detections caused by bone artifacts; and smoothing the boundary of the corrected probability map according to neighborhood consistency to reduce isolated artifacts in the bone's adjacent region.

6. A method for lesion segmentation in medical images for radiology according to claim 1, characterized in that, The total loss function of the brain hemorrhage segmentation network includes: L=L Dice +L CE +λ∑(BoneMask×P ICH ); BoneMask is the bone mask, P ICH This is a bleeding probability plot, where λ is the weighting coefficient.

7. A method for lesion segmentation in medical images for radiology according to claim 1, characterized in that, The multi-channel input images are fused by the channel attention module; the channel attention module generates channel weights based on the response differences of brain windows, subdural windows and bone windows to enhance the feature complementarity between different CT window positions; and enhances the joint expression effect of local structures based on the cross-channel interaction mechanism.

8. A method for lesion segmentation in medical images for radiology according to claim 1, characterized in that, The brain hemorrhage volume, centroid coordinates, diffusion direction index, and symmetry offset index are calculated based on the brain hemorrhage segmentation mask; and structured auxiliary diagnostic information containing lesion range, spatial location, and quantitative indicators is generated based on the above parameters.

9. A lesion segmentation system for medical images used in radiology, characterized in that, The system is used to implement a method for lesion segmentation in medical images for radiology as described in any one of claims 1-8, comprising: Image input module: used to parse DICOM image sequences and perform pixel reconstruction and normalization processing; Multi-window construction module: used to generate brain window, subdural window and bone window images to form a multi-channel input image; Bone Prior Generation Module: Used to generate bone masks and bone boundary probability maps through a bone tissue segmentation network; The brain hemorrhage segmentation module is used to perform encoding feature extraction, feature suppression processing, boundary enhancement processing, and decoding inference to generate an initial hemorrhage probability map. Probability correction module: used to correct the initial hemorrhage probability map based on the bone mask in order to suppress false detections of bone regions; Post-processing module: used for connected component analysis, artifact filtering, and enhancement of small lesions; Clinical parameter calculation module: used to calculate the volume, centroid, and diffusion index of cerebral hemorrhage based on the final segmentation mask and generate auxiliary diagnostic information.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1 to 9.

Images