Brain tumor detection method based on improved YOLO11

Abstract

The invention discloses a brain tumor detection method based on improved YOLO11, which aims at special properties of medical images and clinical requirements of brain tumor detection, and comprises the following steps: firstly, standardizing and dividing brain tumor medical image data; the method comprises the following steps: designing a multi-stage mixed expansion convolution residual module (BTSA-MFE), and embedding the BTSA-MFE modules in a backbone network and a neck network; designing a lightweight medical feature fusion module M IAFA, and embedding an MIAFA module in a backbone network; a medical image specialized dynamic adaptive internal intersection-to-union ratio loss module MDIOU is designed, bounding box regression is adaptively optimized through a tissue density sensing type dynamic proportion adjustment mechanism, and the positioning precision of tumor targets with fuzzy edges and irregular shapes is improved; after training is completed, the brain tumor medical image to be detected is input into the model, the detection result is output, and the detection precision is improved.

Description

Technical Field

[0001] The present invention belongs to the intersection of computer vision and medical image processing, and specifically relates to a target detection method based on improved YOLO11 targeting the characteristics of brain tumor medical images. Background Art

[0002] Brain tumors are common diseases of the central nervous system, and their early and accurate detection plays a decisive role in treatment effectiveness. Unlike common object detection, brain tumors present unique challenges in medical imaging: tumor boundaries are fuzzy and irregular, with unclear transitions from surrounding normal brain tissue; tumors are highly heterogeneous, with complex variations in density, texture, and brightness; and tumor morphology varies greatly, ranging from a few millimeters to several centimeters, with variable distribution.

[0003] Currently, the YOLO family of algorithms is widely used in the field of object detection due to its single-stage detection architecture and real-time performance. However, in the context of brain tumor detection, the standard YOLO11 general-purpose object detection algorithm, applied to medical images, lacks a specific design tailored to the characteristics of brain tumors. This results in insufficient accuracy in clinical applications, particularly for small tumors with blurred boundaries, making it difficult to meet the high-precision requirements of medical diagnosis. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention proposes a brain tumor detection method based on the medical image-specific improved YOLO11. By introducing a dedicated medical image processing module and optimization strategy, the detection effect and clinical practicality of brain tumor medical images are improved.

[0005] In order to achieve the above effects, the present invention adopts the following technical solutions:

[0006] A brain tumor detection method based on improved YOLO11 includes the following steps:

[0007] Step 1: The brain tumor medical imaging dataset is preprocessed according to the medical imaging standardization protocol and divided into training set, validation set and test set;

[0008] Step 2: Build a brain tumor detection model based on the improved YOLO11. The specific process is as follows:

[0009] Step 2.1: Design a multi-level hybrid dilated convolution residual module BTSA-MFE to improve the backbone and neck structures of YOLO11. The BTSA-MFE module consists of a parallel hybrid dilated convolution unit, a multi-level hybrid feature enhancement unit, and a medical image-aware adaptive residual connection unit.

[0010] The parallel hybrid dilated convolution unit includes three parallel hybrid dilated convolution branches, whose dilation rate is set in a nonlinear increasing manner. This increasing mode is specifically designed for the irregular edge morphology characteristics of brain tumors and effectively captures the gradual characteristics of tumor edges. Each branch generates a feature map with different receptive fields of view through a parameterized dilated convolution operation.

[0011] The multi-level hybrid feature enhancement unit is composed of a multi-level hybrid attention module BSPRA. The BSPRA module first adopts a staggered hybrid pooling strategy, which is specifically optimized for the subtle contrast differences between tumors and normal tissues in medical images, enhancing the perception of the unique texture features of brain tumors. The BSPRA module then concatenates the pooling results along the first dimension and, through two layers of nonlinear transformation, generates tumor region-sensitive attention weights to achieve feature enhancement of important brain tumor regions.

[0012] The medical image-aware adaptive residual connection unit first performs dimension alignment, and the final output is the adaptive superposition result of the sum of the weighted features of each branch and the original feature map;

[0013] Step 2.2: Design a lightweight medical feature fusion module MIAFA to improve the backbone structure of YOLO11. The MIAFA module consists of a medical-specific multi-scale feature extraction layer and a brain tumor-specific feature fusion unit.

[0014] The medically specialized multi-scale extraction layer extracts feature maps through three mutually orthogonal convolutions. The first convolution uses a standard convolution as the basic perception layer, the second convolution layer adopts the BTSA-MFE module, and the third convolution layer uses a BTSA-MFE module with a larger dilation rate combination. This special design effectively captures the edge gradients and internal uneven density characteristics unique to brain tumors.

[0015] The brain tumor-specific feature fusion unit splices the feature maps output by the three channels of the medical-specific multi-scale extraction layer, and then enhances the features of important brain tumor areas through the multi-level hybrid attention module BSPRA to generate optimized fusion features;

[0016] Step 2.3: Design a medical imaging-specific dynamic adaptive internal intersection-over-union loss module (MDIOU). This module provides a precise solution to the fuzzy boundary and tissue heterogeneity problems unique to medical imaging by introducing a tissue density-aware dynamic internal self-calibration mechanism and a tumor morphological feature distribution factor (Dfactor). It achieves accurate modeling of tumor directional heterogeneity, scale differences, and tissue overlap quality. While maintaining the advantages of geometric constraints, it automatically adjusts the bounding box positioning weight through pixel-level tumor morphological feature analysis, thereby improving the detection rate of small lesions smaller than 5 mm, the recognition accuracy of irregularly shaped tumors, and the accuracy of edge target detection. The implementation formula of MDIOU is:

[0017]

[0018] Where InnerIoU is the ratio of the internal overlapping area, ρ 2 is the square of the Euclidean distance to the center point, c 2 is the minimum bounding box diagonal square, α is the adaptive penalty weight, v is the aspect ratio consistency measurement parameter, D factor is the tumor density perception factor, which is used to dynamically adjust the loss weight according to the tissue characteristics of the target area. factor The implementation formula is:

[0019]

[0020] Where γ is the scale parameter, θ is the angle difference between frame 1 and frame 2, σ is the normalization parameter, Δs is the scale difference between the two frames, and s max is the largest of the two scales, q overlap is the overlap quality factor;

[0021] Step 3: Input the brain tumor dataset into the improved YOLO11-based brain tumor detection model in step 2 for training. Improve the model performance by optimizing the parameters and save the optimized configuration as the best.pt file.

[0022] Step 4: Process brain tumor imaging data through the optimal brain tumor detection model based on improved YOLO11 to generate tumor lesion detection results.

[0023] Furthermore, the medical imaging data used in step 1 is derived from the open access Br35h medical image database, which covers a single pathological type, namely intracranial tumor samples, and the data is divided into 70% training samples, 20% validation samples, and 10% test samples.

[0024] Furthermore, the optimal parameters of step 3 are set in detail as follows: Adam is used as the optimization method, the training cycle is set to 150 rounds, the batch-size is set to 32, the initial learning rate is set to 0.001, the weight decay parameter is 0.0005, the momentum factor is set to 0.937, and the number of threads is 10.

[0025] Compared with the prior art, the present invention has the following technical effects:

[0026] The present invention uses a multi-level hybrid dilated convolution residual module BTSA-MFE and a brain tumor-specific nonlinear increasing expansion rate combination to effectively capture tumor features of different scales. Compared with the YOLO11 detection method, the detection accuracy of small-sized, fuzzy-edge, and irregular-shaped brain tumor targets is significantly improved. The designed lightweight medical feature fusion module MIAFA combines bidirectional cross-level feature interaction with deformable convolution alignment technology to effectively solve the problem of insufficient fusion of deep and shallow semantic information in the existing technology, and significantly improves the recognition ability of brain tumor targets of different sizes. The MDIOU medical imaging-specific dynamic adaptive loss function uses a tissue density-aware dynamic scale adjustment mechanism to adaptively optimize the special density distribution characteristics of brain tissue. Compared with the traditional CIOU loss function, it has significantly improved the bounding box regression accuracy of fuzzy-edge tumors.

[0027] The experimental results of this invention on the public dataset Br35h show that it performs well in detecting brain tumor medical images, surpassing the original YOLO11 model in terms of recall rate, mAP50, and mAP50-95, and can meet the needs of brain tumor medical image detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is the distribution diagram of each category of targets in the Br35h dataset of the present invention;

[0029] Figure 2 This is a flowchart of a brain tumor detection method based on improved YOLO11 of the present invention;

[0030] Figure 3 Schematic diagram of the structure of a brain tumor detection network model based on improved YOLO11 of the present invention;

[0031] Figure 4 Schematic diagram of the structure of the BTSA-MFE module of the present invention;

[0032] Figure 5 Schematic diagram of the structure of the BSPRA module of the present invention;

[0033] Figure 6 Schematic diagram of the structure of the MIAFA module of the present invention;

[0034] Figure 7(a)-Figure 7(d) These are all partial detection result diagrams of the present invention on the test set. DETAILED DESCRIPTION

[0035] The specific contents of the present invention are further described in detail below with reference to implementation cases.

[0036] A brain tumor detection method based on improved YOLO11 includes the following steps:

[0037] Step 1: The brain tumor medical imaging dataset is preprocessed according to the medical imaging standardization protocol and divided into training set, validation set and test set;

[0038] The brain tumor medical imaging dataset is derived from the open access Br35h medical image library. The database covers a single pathological type, namely intracranial tumor samples. The data is divided into 70% training samples, 20% validation samples, and 10% test samples.

[0039] Step 2: Build Figure 3 The brain tumor detection model based on the improved YOLO11 shown in the figure has the following specific process:

[0040] Step 2.1: Design Figure 4 The multi-level hybrid dilated convolution residual module BTSA-MFE shown is used to improve the backbone and neck structures of YOLO11, where the BTSA-MFE module consists of a parallel hybrid dilated convolution unit, a multi-level hybrid feature enhancement unit, and a medical image-aware adaptive residual connection unit;

[0041] like Figure 4 As shown in the figure, the parallel hybrid dilated convolution unit contains three parallel hybrid dilated convolution branches, whose dilation rate is set in a nonlinear increasing manner. This increasing mode is specifically designed for the irregular edge morphology characteristics of brain tumors and effectively captures the gradual characteristics of tumor edges. Each branch generates a feature map with different receptive fields of view through parameterized dilated convolution operations;

[0042] like Figure 5 As shown in the figure, the multi-level hybrid feature enhancement unit is composed of a multi-level hybrid attention module BSPRA. The BSPRA module first adopts a staggered hybrid pooling strategy, which is specially optimized for the subtle contrast difference between tumors and normal tissues in medical images, enhancing the perception of the unique texture features of brain tumors. Then, the BSPRA module splices the pooling results along the first dimension and generates tumor region-sensitive attention weights through two layers of nonlinear transformation to achieve feature enhancement of important brain tumor areas.

[0043] like Figure 5As shown, the medical image-aware adaptive residual connection unit first performs dimension alignment, and the final output is the adaptive superposition result of the sum of the weighted features of each branch and the original feature map;

[0044] Step 2.2: Design Figure 6 The lightweight medical feature fusion module MIAFA is used to improve the backbone structure of YOLO11. The MIAFA module consists of a medical-specific multi-scale feature extraction layer and a brain tumor-specific feature fusion unit.

[0045] like Figure 6 As shown in the figure, the medical-specific multi-scale extraction layer extracts feature map features through three mutually orthogonal convolutions. The first convolution uses standard convolution as the basic perception layer, the second convolution layer adopts the BTSA-MFE module, and the third convolution layer uses the BTSA-MFE module with a larger dilation rate combination. This special design effectively captures the edge gradient and internal uneven density characteristics unique to brain tumors.

[0046] like Figure 6 As shown in the figure, the brain tumor-specific feature fusion unit splices the feature maps output by the three-way medical-specific multi-scale extraction layer, and then enhances the features of important brain tumor areas through the multi-level hybrid attention module BSPRA to generate optimized fusion features;

[0047] Step 2.3: Design a medical imaging-specific dynamic adaptive internal intersection-over-union loss module (MDIOU). This module provides a precise solution to the fuzzy boundary and tissue heterogeneity problems unique to medical imaging by introducing a tissue density-aware dynamic internal self-calibration mechanism and a tumor morphological feature distribution factor (Dfactor). It achieves accurate modeling of tumor directional heterogeneity, scale differences, and tissue overlap quality. While maintaining the advantages of geometric constraints, it automatically adjusts the bounding box positioning weight through pixel-level tumor morphological feature analysis, thereby improving the detection rate of small lesions smaller than 5 mm, the recognition accuracy of irregularly shaped tumors, and the accuracy of edge target detection. The implementation formula of MDIOU is:

[0048]

[0049] Where InnerIoU is the ratio of the internal overlapping area, ρ 2 is the square of the Euclidean distance to the center point, c 2 is the minimum bounding box diagonal square, α is the adaptive penalty weight, v is the aspect ratio consistency measurement parameter, D factor is the tumor density perception factor, which is used to dynamically adjust the loss weight according to the tissue characteristics of the target area. factor The implementation formula is:

[0050]

[0051] Where γ is the scale parameter, θ is the angle difference between frame 1 and frame 2, σ is the normalization parameter, Δs is the scale difference between the two frames, and s max is the largest of the two scales, q overlap is the overlap quality factor;

[0052] Step 3: Input the brain tumor dataset into the improved YOLO11-based brain tumor detection model in step 2 for training. Improve the model performance by optimizing the parameters and save the optimized configuration as the best.pt file.

[0053] The parameter setting process is as follows: Adam is used as the optimization method, and the training cycle is set to 150 rounds, the batch-size is set to 32, the initial learning rate is set to 0.001, the weight decay parameter is set to 0.0005, the momentum factor is set to 0.937, and the number of threads is 10. The specific experimental environment is shown in Table 1.

[0054] Table 1 Experimental environment parameters

[0055]

[0056] Step 4: Process brain tumor imaging data using the optimal brain tumor detection model based on improved YOLO11 to generate tumor lesion detection results. The specific process is as follows:

[0057] The brain tumor medical imaging data to be detected is input into the optimal brain tumor detection model based on improved YOLO11 saved in step 3 for detection, and the output is as follows: Figures 7(a) to 7(d) The images shown are labeled with the detected object categories and their corresponding confidence scores, from Figures 7(a) to 7(d) It can be seen that the brain tumor detection model based on the improved YOLO11 can accurately detect the target image of brain tumors.

[0058] To verify the improved detection accuracy of the brain tumor detection model based on the improved YOLO11 proposed in this example, Recall, mAP50 (average precision at an intersection-over-union threshold of 0.5), and mAP50-90 (average precision with intersection-over-union thresholds ranging from 0.5 to 0.95) were used as key indicators to evaluate the detection results. Table 2 shows a model performance comparison table. Under consistent experimental environments and parameters, the recall value of the proposed algorithm reached 92.1%, an improvement of 1.9% compared to YOLO11. The mAP50 value reached 95.8%, an improvement of 0.8% compared to YOLO11. The mAP50-90 value reached 71.7%, an improvement of 1.4% compared to YOLO11. This proves that the brain tumor detection method based on the improved YOLO11 proposed in this article is feasible.

[0059] Table 2 Performance comparison table

[0060]

Claims

1. A brain tumor detection method based on improved YOLO11, characterized in that: The steps include: Step 1: The brain tumor medical imaging dataset is preprocessed according to the medical imaging standardization protocol and divided into training set, validation set and test set; Step 2: Build a brain tumor detection model based on the improved YOLO11. The specific process is as follows: Step 2.1: Design a multi-level hybrid dilated convolution residual module BTSA-MFE to improve the backbone and neck structures of YOLO11. The BTSA-MFE module consists of a parallel hybrid dilated convolution unit, a multi-level hybrid feature enhancement unit, and a medical image-aware adaptive residual connection unit. The parallel hybrid dilated convolution unit is a parallel hybrid dilated convolution branch, and its dilation rate is set in a nonlinear increasing manner. This increasing mode is specifically designed for the irregular edge morphology characteristics of brain tumors, effectively capturing the gradual characteristics of tumor edges. Each branch generates a feature map with different receptive fields of view through a parameterized dilated convolution operation; The multi-level hybrid feature enhancement unit is composed of a multi-level hybrid attention module BSPRA. The BSPRA module first adopts a staggered hybrid pooling strategy, which is specifically optimized for the subtle contrast differences between tumors and normal tissues in medical images, enhancing the perception of the unique texture features of brain tumors. The BSPRA module then concatenates the pooling results along the first dimension and, through two layers of nonlinear transformation, generates tumor region-sensitive attention weights to achieve feature enhancement of important brain tumor regions. The medical image-aware adaptive residual connection unit first performs dimension alignment, and the final output is the adaptive superposition result of the sum of the weighted features of each branch and the original feature map; Step 2.2: Design a lightweight medical feature fusion module MIAFA to improve the backbone structure of YOLO11. The MIAFA module consists of a medical-specific multi-scale feature extraction layer and a brain tumor-specific feature fusion unit. The medically specialized multi-scale extraction layer extracts feature maps through mutually orthogonal convolutions. The first convolution layer uses standard convolution as the basic perception layer, the second convolution layer adopts the BTSA-MFE module, and the third convolution layer uses the BTSA-MFE module with a larger nonlinear incremental expansion rate. This special design effectively captures the edge gradient and internal uneven density characteristics unique to brain tumors. The brain tumor-specific feature fusion unit splices the feature maps output by the medical-specific multi-scale extraction layer, and then enhances the features of important brain tumor areas through the multi-level hybrid attention module BSPRA to generate optimized fusion features; Step 2.3: Design a medical imaging-specific dynamic adaptive internal intersection-over-union loss module (MDIOU). This module provides a precise solution to the fuzzy boundary and tissue heterogeneity problems unique to medical imaging by introducing a tissue density-aware dynamic internal self-calibration mechanism and a tumor morphological feature distribution factor (Dfactor). It achieves accurate modeling of tumor directional heterogeneity, scale differences, and tissue overlap quality. While maintaining the advantages of geometric constraints, it automatically adjusts the bounding box positioning weight through pixel-level tumor morphological feature analysis, thereby improving the detection rate of small lesions smaller than 5 mm, the recognition accuracy of irregularly shaped tumors, and the accuracy of edge target detection. The implementation formula of MDIOU is: Where InnerIoU is the ratio of the internal overlapping area, ρ 2 is the square of the Euclidean distance to the center point, c 2 is the minimum bounding box diagonal square, α is the adaptive penalty weight, v is the aspect ratio consistency measurement parameter, D factor is the tumor density perception factor, which is used to dynamically adjust the loss weight according to the tissue characteristics of the target area. factor The implementation formula is: Where γ is the scale parameter, θ is the angle difference between frame 1 and frame 2, σ is the normalization parameter, Δs is the scale difference between the two frames, and s max is the largest of the two scales, q overlap is the overlap quality factor. Step 3: Input the brain tumor dataset into the improved YOLO11-based brain tumor detection model in step 2 for training. Improve the model performance by optimizing the parameters and save the optimized configuration as the best.pt file. Step 4: Process brain tumor imaging data through the optimal brain tumor detection model based on improved YOLO11 to generate tumor lesion detection results.

2. The brain tumor detection method based on improved YOLO11 according to claim 1, characterized in that: The medical imaging data used in step 1 is from the open access Br35h medical image database, which covers a single pathological type, namely intracranial tumor samples.

Images