Non-functional pancreatic neuroendocrine tumor focus segmentation method and system

Abstract

The invention belongs to the technical field of medical image processing, and particularly relates to a non-functional pancreatic neuroendocrine tumor focus segmentation method and system. According to the method, an improved U-Net architecture is combined with position embedding (PE) and symbol distance field (SDF) supervision strategy, and high-precision and high-efficiency NF-PNETs focus segmentation is realized by fusing multi-scale feature extraction, lightweight 2.5 D convolution design and a human-participated circular labeling process; and meanwhile, through an AI auxiliary labeling and iterative training mechanism, 60% of manual labeling time is reduced. The method shows strong robustness for complex lesions and multi-protocol CT images, can effectively support preoperative accurate diagnosis and personalized treatment planning, and has wide clinical application value.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to a non-functional pancreatic neuroendocrine tumor lesion segmentation method and system. BACKGROUND

[0002] Preoperative accurate segmentation of non-functional pancreatic neuroendocrine tumors (NF-PNETs) is crucial for clinical decision-making, directly affecting surgical planning, lymph node dissection range and prognosis evaluation. Traditional methods mainly rely on manual annotation by radiologists or automatic segmentation models based on convolutional neural networks (CNN) and Transformers, but still have significant limitations, such as insufficient segmentation accuracy, which cannot meet the clinical demand for high-precision lesion contours; low annotation efficiency, full manual annotation is time-consuming and costly, and single instance annotation time usually exceeds 30 minutes, and is easily affected by subjective judgment, making it difficult to support the construction of large-scale, multi-center datasets; limited generalization ability, differences in CT imaging protocols (such as slice thickness and contrast agent injection schemes) between different medical centers result in significant performance degradation of the model on external datasets; fuzzy boundaries and artifact interference, complex anatomical structures around the pancreas, partial volume effects and motion artifacts easily cause segmentation results to be broken or overly smoothed, affecting the reliability of downstream pathological grading and metastasis risk assessment.

[0003] Although existing research has partially improved performance by introducing a three-dimensional U-Net or hybrid Transformer-CNN architecture, the high computational complexity makes it difficult to adapt to clinical hardware resources. In addition, the purely automated annotation process lacks human expert correction mechanisms, leading to error accumulation and model bias.

[0004] Therefore, there is an urgent need for an efficient, high-precision and robust automatic segmentation solution that can simultaneously address complex lesion segmentation, cross-protocol data adaptation and low annotation efficiency, providing reliable support for preoperative accurate diagnosis and individualized treatment. SUMMARY

[0005] The purpose of the present application is to provide a non-functional pancreatic neuroendocrine tumor lesion segmentation method and system that can efficiently and accurately segment complex lesions, while addressing the problems of insufficient segmentation accuracy, low annotation efficiency and poor cross-center generalization ability in existing technology.

[0006] The non-functional pancreatic neuroendocrine tumor lesion segmentation method provided by the present application includes the following specific steps:

[0007] (1) Preprocess the input abdominal CT image for multi-protocol adaptation;

[0008] (2) Using the TotalSegmentator tool to automatically locate the pancreatic region;

[0009] (3) Based on the improved U-Net model for fine segmentation of the located pancreatic region, the U-Net model uses position embedding (PE) strategy;

[0010] (4) The model uses signed distance field (SDF) supervision, and the model training adopts a composite loss function to optimize the segmentation;

[0011] (5) Through a human-involved iterative annotation process, the segmentation model is iteratively optimized, specifically including AI-generated initial segmentation results, interactive correction by radiologists, and feedback of the corrected data to the model training.

[0012] Further:

[0013] In step (1), the input abdominal CT image is preprocessed by multi-protocol adaptation, specifically: first, the input CT image C i is isotropically resampled (resolution 0.7x0.7x1mm 3 ) and intensity normalized; the standardized display parameters of window width 250HU and window level 50HU are used to enhance the visual contrast of the lesion area, and the preprocessed image I i is obtained.

[0014] In step (2), the TotalSegmentator tool is used to automatically locate the pancreatic region, generate the pancreatic region of interest, crop and scale to a fixed size (384x384x64 pixels), and obtain the ROI region R i .

[0015] In step (3), the improved U-Net model is specifically: the traditional 3D convolution is decomposed into axial 2D plane convolution (e.g., kernel 5x5) and depth 1D convolution (e.g., kernel 5), reducing the computational complexity by 76% (arithmetic operations from 125 times per voxel to 30 times per voxel); the single slice S i of the input image I i is injected with learnable position encoding (PE) along the z-axis (depth direction) to distinguish adjacent slices;

[0016]

[0017] wherein z is the slice position and D is the encoding dimension.

[0018] The PE encoding is added to the input slice channel by channel to enhance the model's perception of the slice order.

[0019] The single slice S iThe feature after PE encoding (size 384x384xC, C is the number of channels). 2D convolution is performed in the x-y plane (e.g., kernel 5x5, stride 2, padding 2) to obtain the output feature map F xy The feature after PE encoding (size 384x384xC, C is the number of channels). 2D convolution is performed in the x-y plane (e.g., kernel 5x5, stride 2, padding 2) to obtain the output feature map F z The feature after PE encoding (size 384x384xC, C is the number of channels). 2D convolution is performed in the x-y plane (e.g., kernel 5x5, stride 2, padding 2) to obtain the output feature map F i The feature after PE encoding (size 384x384xC, C is the number of channels). 2D convolution is performed in the x-y plane (e.g., kernel 5x5, stride 2, padding 2) to obtain the output feature map F

[0020] The feature after PE encoding (size 384x384xC, C is the number of channels). 2D convolution is performed in the x-y plane (e.g., kernel 5x5, stride 2, padding 2) to obtain the output feature map F i Up-sampling is performed using a 2x2 kernel to restore the resolution to 48x48x64. The feature F3 (48x48x64x256) of the third layer of the encoder is spliced along the channel dimension to generate a feature map of (48x48x64x(256+512)). Two groups of convolution operations are performed on the spliced feature, and the output channel number is reduced to 256. The F3, F2, and F1 features are sequentially up-sampled and spliced, and finally restored to the initial resolution (384x384x64).

[0021] The model described in step (4) is supervised using a signed distance field (SDF), which specifically includes: using the output (384x384x64x64) of the last layer of the decoder. The channel number is compressed to 1 to generate the original segmentation mask M raw The feature after PE encoding (size 384x384xC, C is the number of channels). 2D convolution is performed in the x-y plane (e.g., kernel 5x5, stride 2, padding 2) to obtain the output feature map F

[0022] The feature after PE encoding (size 384x384xC, C is the number of channels). 2D convolution is performed in the x-y plane (e.g., kernel 5x5, stride 2, padding 2) to obtain the output feature map F raw The feature after PE encoding (size 384x384xC, C is the number of channels). 2D convolution is performed in the x-y plane (e.g., kernel 5x5, stride 2, padding 2) to obtain the output feature map F pred

[0023]

[0024] wherein, is the mask boundary, and the distance calculation uses the Fast Marching Method.

[0025] The difference between the predicted and real SDFs is optimized by L1 loss:

[0026]

[0027] wherein, SDF pred and SDF gt are the predicted and real signed distance fields, respectively.

[0028] ​The model training in step (4) adopts a composite loss function, including a Dice loss for optimizing the overlap degree of the lesion area segmentation, an SDF constraint L1 loss for refining the boundary smoothness, and a classification cross-entropy loss for joint training of the auxiliary task:

[0029]

[0030] Wherein, λ1, λ2, λ3 are weight coefficients, balancing the optimization objectives of different tasks.

[0031] The human-involved cyclic annotation process in step (5) specifically includes: generating an initial segmentation result by using a pre-trained model, and performing interactive boundary correction through a 3D Slicer plug-in; adding the corrected annotation data to the training set, and forming a "annotation-training-optimization" closed loop through iterative training; the single annotation time is reduced by 60%, and dynamic expansion of cross-center data sets is supported.

[0032] The application also provides a non-functional pancreatic neuroendocrine tumor lesion segmentation system based on the above segmentation method, which comprises: a data preprocessing module for CT image standardization, resampling and ROI extraction; an improved U-Net segmentation module for executing the segmentation method; and an AI-assisted annotation platform integrated with a 3D Slicer tool to support interactive annotation and model iterative optimization.

[0033] Further, the AI-assisted annotation platform automatically identifies error-prone samples and prioritizes pushing them to radiologists for correction, and supports quality evaluation and abnormal value filtering of batch annotation results to improve efficiency.

[0034] The application also includes a computer-readable storage medium storing program code for executing the non-functional pancreatic neuroendocrine tumor lesion segmentation method.

[0035] The application breaks through the limitations of traditional segmentation models in terms of insufficient precision, low annotation efficiency and poor generalization, and innovatively optimizes key links such as pancreatic region positioning, lesion boundary refinement and cross-center data adaptation, significantly improving segmentation performance, while reducing 60% of the manual annotation time through AI-assisted annotation and iterative training mechanism. The application shows strong robustness to complex lesions and multi-protocol CT images, and can effectively support preoperative accurate diagnosis and individualized treatment planning, and has wide clinical application value.

[0036] Compared with the prior art, the improved U-Net architecture combined with position embedding (PE) and signed distance field (SDF) supervision strategy, the fusion of multi-scale feature extraction, lightweight 2.5D convolution design and human-involved recurrent annotation process, solves the core pain points of insufficient precision, complex calculation and poor generalization of existing segmentation methods, provides an efficient and reliable AI tool for precise diagnosis and treatment of non-functional pancreatic neuroendocrine tumors, and has significant clinical application value and commercialization potential. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flow chart of the non-functional pancreatic neuroendocrine tumor lesion segmentation method of an embodiment of the present application.

[0038] Figure 2 Qualitative comparison of the PNET segmentation of the method of the present application and other benchmark deep learning methods on different data sets. Original CT images, overlaid ground truth segmentation (purple), predictions from different models, and 3D renderings of the segmentation results (red for model predictions, green for ground truth). DETAILED DESCRIPTION

[0039] The present application will be described in detail below with specific embodiments. The following examples will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made. These all belong to the protection scope of the present application.

[0040] Specifically, the present application provides a non-functional pancreatic neuroendocrine tumor lesion segmentation method, as shown in Figure 1 The specific steps are:

[0041] First, the input CT image C i is isotropically resampled (resolution 0.7x0.7x1mm 3 ) and intensity normalized; the standardized display parameters of window width 250HU and window level 50HU are used to enhance the visual contrast of the lesion area, and the preprocessed image I norm is obtained:

[0042] μ liver = 50HU, σ liver = 25HU

[0043] The pancreatic region is automatically located using the TotalSegmentator tool, the pancreatic region of interest is generated, and the ROI region R i is cropped and scaled to a fixed size (384x384x64 pixels).

[0044] The traditional 3D convolution is decomposed into axial 2D plane convolution (kernel 5x5) and depth 1D convolution (kernel 5), reducing the computation by 76% (arithmetic operations from 125 per voxel to 30 per voxel). The input image I i is split into single slices S i A learnable positional encoding (PE) is injected along the z-axis (depth direction) to distinguish adjacent slices;

[0045] or

[0046] where z is the slice position and D is the encoding dimension.

[0047] The PE encoding is added to the input slice channel by channel, enhancing the model's perception of slice order.

[0048] The single slice S i is concatenated with the PE-encoded features (size 384x384xC, C is the number of channels). 2D convolution (kernel 5x5, stride 2, padding 2) is performed in the x-y plane to obtain the output feature map F xy (size 192x192xC1). 1D convolution (kernel 5, stride 1, padding 2) is performed on adjacent 3 slices along the z-axis to obtain the output feature map F z (size 192x192xC1).

[0049] Downsampling is repeated 4 times (resolution halved each time), and each layer structure contains 2 groups of "2.5D convolution + batch normalization + ReLU activation". 2x2 pooling is performed in the x-y plane (only the first 3 layers) to obtain the feature map F i and each layer feature F3, F2, F1.

[0050] The obtained feature map F i is upsampled using a 2x2 kernel to restore the resolution to 48x48x64. It is concatenated with the feature F3 (48x48x64x256) of the 3rd layer of the encoder along the channel dimension to generate a feature map of (48x48x64x(256+512)). Perform 2 groups of convolution operations on the concatenated features, and reduce the output channel number to 256. In turn, upsample and concatenate F3, F2, F1 features to finally restore to the initial resolution (384x384x64).

[0051] The output of the last layer of the decoder (384x384x64x64) is used. The channel number is compressed to 1 to generate the original segmentation mask M raw (size 384x384x64x1). Convert M raw to signed distance field SDF pred :

[0052]

[0053] in, To mask the boundary, the distance calculation uses the Fast Marching Method.

[0054] The difference between the predicted and true SDF is optimized by L1 loss:

[0055]

[0056] Among them, SDFD pred and SDF gt are the predicted and true signed distance fields, respectively.

[0057] The model training adopts a composite loss function, including Dice loss for optimizing the overlap of lesion region segmentation, SDF constrained L1 loss for refining boundary smoothness, and classification cross entropy loss for joint training of auxiliary tasks.

[0058]

[0059] Among them, λ1, λ2, and λ3 are weight coefficients that balance the optimization objectives of different tasks.

[0060] The pre-trained model is used to generate initial segmentation results and interactive boundary correction is performed through the 3D Slicer plug-in. The corrected labeled data is added to the training set, and a "labeling-training-optimization" closed loop is formed through iterative training. The single-case labeling time is reduced by 60%, and dynamic expansion of cross-center data sets is supported.

[0061] The qualitative performance of the proposed method is compared with other benchmark deep learning methods (including 2D U-Net, UNETR, and TransUNet) for pancreatic neuroendocrine tumor (PNET) segmentation on different datasets. Figure 2As shown, five representative sample types are covered, from the training set, test set, internal validation set, and external data from Beijing and Guangzhou. Among them, sample a is a typical medium-sized tumor with clear boundaries. The method of the application can accurately retain the irregular structure of the tumor surface, which is better than other methods with blurred boundaries or false positives; sample b is a small tumor with fuzzy edges. The method of the application realizes accurate positioning under the constraint of SDF boundary, while other methods have obvious missed detection or excessive expansion phenomenon; sample c is a tumor with heterogeneous internal structure. The method of the application successfully distinguishes between necrotic and active areas, while the benchmark model is prone to misjudgment; sample d is collected from a 5mm-thick CT. Traditional methods are disturbed by interlayer artifacts and show "stepped" fractures, while the method of the application eliminates this problem through isotropic resampling and PE position coding, significantly enhancing segmentation continuity; sample e is a large volume tumor with complex morphology. The method of the application completely restores its three-dimensional structure and lobular details, which is better than other models with holes or topological errors.

[0062] Further, the 3D reconstruction result shows that the tumor model generated by the method of the application has smooth structure and reasonable anatomy, with a volume error controlled within 5%, which is much better than the rough and incomplete reconstruction results of other models. The PE position coding (D=64) introduced in the application effectively deals with the interlayer discontinuity problem, the SDF boundary supervision (λ2=0.3) significantly reduces the error distance (HD95 reduces to 34.1mm) of the fuzzy edge, and the 2.5D convolution structure realizes 76% compression of the calculation amount with 5x5 axial and depth kernel, which maintains the performance stability while ensuring high-resolution input processing. Overall, the method of the application significantly improves the Dice coefficient in various challenging scenarios (an average increase of 8.8%, 0.74 vs. 0.68), with a cross-center generalization error fluctuation controlled within 10%, showing good robustness and clinical application potential, especially in the spatial structure reconstruction of three-dimensional surgical planning, which has important application value, fully embodying the comprehensive innovation ability of the method of the application in position perception, boundary sensitivity and efficient modeling.

[0063] The specific embodiments of the application are described above. It should be understood that the application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the application. The embodiments of the present application and the features in the embodiments can be arbitrarily combined without conflict.

Claims

1. A method for segmenting non-functional pancreatic neuroendocrine tumor lesions, characterized in that: The specific steps are: (1) Perform multi-protocol adaptation preprocessing on the input abdominal CT image; (2) automatic localization of the pancreatic region using the TotalSegmentator tool; (3) performing refined lesion segmentation on the located pancreatic region based on an improved U-Net model using a position embedding (PE) strategy; (4) The model is supervised by signed distance field (SDF), and the model training adopts composite loss function to jointly optimize the segmentation; (5) The segmentation model is iteratively optimized through a cyclic labeling process involving human participation, which includes AI generating initial segmentation results, interactive corrections by radiologists, and feedback of the corrected data to model training.

2. The method according to claim 1, characterized in that The multi-protocol adaptation preprocessing of the input abdominal CT image in step (1) is as follows: first, the input CT image C i Perform isotropic resampling and intensity normalization; The standardized display parameters of window width 250HU and window level 50HU were used to enhance the visual contrast of the lesion area and obtain the preprocessed image I. i .

3. The method according to claim 2, characterized in that As described in step (2), the TotalSegmentator tool is used to automatically locate the pancreatic region, generate the pancreatic region of interest, crop and scale it to a fixed size, and obtain the ROI area R i .

4. The method according to claim 3, characterized in that The improved U-Net model described in step (3) specifically decomposes the traditional 3D convolution into axial 2D plane convolution and depth 1D convolution; for the input image I i A single slice S i Injecting learnable positional encodings (PEs) along the z-axis to distinguish adjacent slices; or Where z is the slice position and D is the encoding dimension; The PE code is added to the input slice channel by channel to enhance the model's awareness of the slice order.

5. The method according to claim 4, characterized in that In step (3), the single slice S i +PE encoded features, single slice S i The size is 384×384×1, and the feature size after PE encoding is 384×384×C, where C is the number of channels; perform 2D convolution on the xy plane to obtain the output feature map F xy ; Perform 1D convolution on three adjacent slices along the z axis to obtain the output feature map F z Repeat 4 downsamplings, each time halving the resolution. Each layer contains 2 sets of "2.5D convolution + batch normalization + ReLU activation". Perform 2×2 pooling in the xy plane. Only the first 3 layers are used to obtain the feature map F. i With each layer feature map F3, F2, F1; Using the obtained feature map F i , upsampling is performed using a 2×2 kernel to restore the resolution to 48×48×64; it is concatenated with the feature F3 (48×48×64×256) of the third layer of the encoder along the channel dimension to generate a feature map of 48×48×64×(256+512); two sets of convolution operations are performed on the concatenated features, and the number of output channels is reduced to 256; the F3, F2, and F1 features are upsampled and concatenated in turn, and finally restored to the initial resolution of 384×384×64.

6. The method according to claim 5, characterized in that The model described in step (4) uses signed distance field (SDF) supervision, specifically including: using the output of the last layer of the decoder (384×384×64×64); compressing the number of channels to 1 to generate the original segmentation mask M raw , the size is 384×384×64×1; raw Convert to signed distance field SDF pred : in, To mask the boundaries, the distance calculation uses the fast marching algorithm; The difference between the predicted and true SDF is optimized by L1 loss: Among them, SDF pred and SDF gt are the predicted and true signed distance fields, respectively.

7. The method according to claim 6, characterized in that The model training in step (4) adopts a composite loss function, including Dice loss for optimizing the overlap of lesion region segmentation, SDF constrained L1 loss for refining boundary smoothness, and classification cross entropy loss for joint training of auxiliary tasks: Among them, λ1, λ2, and λ3 are weight coefficients that balance the optimization objectives of different tasks.

8. The method according to claim 7, characterized in that The human-involved cyclic labeling process described in step (5) specifically includes: using the pre-trained model to generate the initial segmentation results, and interactively correcting the boundaries through the 3D Slicer plug-in; adding the corrected labeled data to the training set, and forming a "labeling-training-optimization" closed loop through iterative training; the single-case labeling time is reduced by 60%, and dynamic expansion of cross-center data sets is supported.

9. A non-functional pancreatic neuroendocrine tumor lesion segmentation system based on the above segmentation method, characterized in that: include: Data preprocessing module for CT image standardization, resampling, and ROI extraction; improved U-Net segmentation module to implement segmentation methods; AI-assisted annotation platform, integrated with 3D Slicer tools to support interactive annotation and model iterative optimization; The AI-assisted labeling platform improves efficiency by automatically identifying error-prone samples and sending them to radiologists for correction, as well as supporting quality assessment and outlier filtering of batch labeling results.

10. A computer-readable storage medium storing program code, wherein the program code is used to execute the method for segmenting non-functional pancreatic neuroendocrine tumor lesions.

Images