PICC (Peripherally Inserted Central Catheter) identification method and system based on image identification

Abstract

The invention discloses a PICC catheter recognition method and system based on image recognition, and relates to the technical field of medical image analysis, and the system comprises a medical image input module which is used for obtaining X-ray image data; the improved YOLO-Seg segmentation module is connected with the medical image input module, and the improved YOLO-Seg segmentation module comprises a CSPDarknet backbone network of which the channel number is expanded to 1280, a feature pyramid network for outputting a 4-scale feature map, and a segmentation head containing 32-channel prototype mask generation; and the double attention enhancement module is integrated in the improved YOLO-Seg segmentation module, comprises a channel attention module and a space attention module, and is used for enhancing low-contrast catheter feature response. According to the method, the slender feature extraction capability is enhanced through the 32-channel prototype mask generator, and a catheter continuity reconstruction algorithm based on skeleton analysis is adopted, so that the Dice coefficient of the PICC catheter is improved, the fracture rate of the catheter is reduced, the problem of slender structure segmentation failure is solved, and the catheter segmentation precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image analysis, in particular to a PICC catheter recognition method and system based on image recognition. BACKGROUND

[0002] PICC catheterization is the insertion of a central venous catheter through a peripheral vein. For patients who need long-term infusion, tumor chemotherapy or critical illness, it is very convenient and timely to infuse drugs for treatment. The chest is composed of a row of ribs, like a row of small fences, protecting the important organs inside. There are some gaps between these small fences, called intercostal spaces. The end of the PICC catheter is located at the level of the 5th and 6th intercostal spaces, because this position is close to the superior vena cava, a large blood vessel in the body responsible for sending blood from the upper half of the body back to the heart. By placing the end of the PICC catheter here, it can ensure that the drug and nutrient solution can smoothly enter the large blood vessel and then be transported to all parts of the body to play a role. To determine that the catheter tip is in the 5th and 6th intercostal spaces, X-ray chest radiography is needed. By taking a chest radiograph, doctors can clearly see the position of the catheter tip and then compare it with the standard position to ensure that the PICC catheter is placed in the right position. The existing technology has the following problems:

[0003] The existing U-Net, Mask R-CNN and other models have significant defects in PICC catheter segmentation. The catheter diameter is only 1-3 pixels, and the traditional segmentation method is prone to breakage. At the same time, the clinical workflow requires >30FPS, while Mask R-CNN only has 18FPS, which cannot meet the real-time intervention requirements, and the model volume reaches 89-245MB, which is difficult to deploy to medical edge devices.

[0004] At the same time, the contrast of the PICC catheter under X-ray is very low, and it is blocked by the rib anatomy, resulting in weak response of standard YOLO-Seg and other general object detection models to catheter features, low segmentation accuracy, and lack of morphological optimization mechanism for slender targets in existing methods, resulting in high catheter mask breakage rate.

[0005] And the existing mainstream solution ignores the characteristics of medical images and does not use the spatial relationship between the ribs and the catheter for consistency verification, resulting in high false positive rate and lack of reconstruction algorithm for catheter continuity such as skeleton analysis, affecting the recognition accuracy. SUMMARY

[0006] The present application provides a PICC catheter recognition method and system based on image recognition to solve the problems raised in the background art.

[0007] To solve the above technical problems, the technical solution adopted by the present application is:

[0008] A PICC catheter identification method based on image recognition includes the following steps:

[0009] S1: Acquire X-ray images and preprocess them to a resolution of 512×512;

[0010] S2: Object detection and instance segmentation are performed simultaneously by improving the YOLO-Seg model, where: a 32-channel prototype mask is used to generate slender duct segmentation results, and a dual attention mechanism is applied to enhance low-contrast region features;

[0011] S3: Intelligent post-processing of the output mask: Reconstructing duct continuity based on skeleton analysis and reducing false positives by utilizing rib spatial position constraints;

[0012] S4: Output the recognition results to the display terminal or clinical navigation system.

[0013] A further improvement of the technical solution of the present invention is that the improved YOLO-Seg model is optimized through multi-task collaborative training: the detection head and the segmentation head are trained together, and an edge detection auxiliary task is added. The loss function includes an anatomical position constraint term, which forces the prediction results to conform to the clinical anatomical distribution.

[0014] A further improvement of the technical solution of the present invention is that: in the intelligent post-processing of anatomy, the continuity reconstruction of the duct adopts pixel connectivity analysis and path search algorithm, and the spatial relationship verification rejects the prediction of ducts located behind the ribs and without fluoroscopic overlap.

[0015] A PICC catheter identification system based on image recognition includes: a medical image input module for acquiring X-ray image data; an improved YOLO-Seg segmentation module connected to the medical image input module, the improved YOLO-Seg segmentation module including: a CSPDarknet backbone network with 1280 channels, a feature pyramid network outputting 4-scale feature maps, and a segmentation head containing 32-channel prototype mask generation; a dual attention enhancement module integrated into the improved YOLO-Seg segmentation module, including a channel attention module and a spatial attention module, for enhancing the response of low-contrast catheter features; and an anatomical intelligent post-processing module for performing rib morphology optimization, catheter skeleton continuity reconstruction, and anatomical spatial relationship verification on the mask output by the segmentation module.

[0016] A further improvement of the technical solution of the present invention is that the loss function of the improved YOLO-Seg segmentation module includes: boundary detection loss, Focal Loss for focusing on slender structures, boundary awareness loss, and anatomical loss based on rib position constraints.

[0017] A further improvement to the technical solution of the present invention is that the intelligent post-anatomical processing module performs the following operations:

[0018] A1: Morphological optimization of the rib mask by sequentially performing opening and closing operations;

[0019] A2: Use skeleton analysis algorithm to connect broken segments on the PICC catheter mask;

[0020] A3: Correcting linear anatomical structures using Hough transform;

[0021] A4: Verify whether the spatial distance between the catheter and the rib is within the preset clinical safety threshold.

[0022] A further improvement of the technical solution of the present invention is that: in the dual attention enhancement module: the channel attention module recalibrates and enhances the response of the catheter region by weighting the feature channels, and the spatial attention module generates a spatial weight map to focus on the anatomical edge.

[0023] A further improvement of the technical solution of the present invention is that it also includes a deployment optimization module, which performs FP16 quantization and layer fusion through TensorRT, supports dynamic input of 512-1024px resolution, and exports ONNX format models to medical edge devices for real-time inference.

[0024] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows:

[0025] 1. This invention provides a PICC catheter identification method and system based on image recognition. It enhances the ability to extract slender features through a 32-channel prototype mask generator and uses a catheter continuity reconstruction algorithm based on skeleton analysis to improve the Dice coefficient of PICC catheters, reduce the catheter breakage rate, solve the problem of failure in slender structure segmentation, and improve catheter segmentation accuracy.

[0026] 2. This invention provides a PICC catheter identification method and system based on image recognition. It reduces the amount of computation by using deep separable convolutions in the CSPDarknet backbone network and optimizes the deployment of TensorRT with FP16 quantization and layer fusion, thereby achieving an inference speed of 38 FPS, which is greater than the clinical requirement of 30 FPS. Moreover, the model size is smaller than that of Mask R-CNN, breaking through the real-time bottleneck and meeting the needs of clinical intervention.

[0027] 3. This invention provides a PICC catheter identification method and system based on image recognition. It enhances the response of low-contrast catheters through a dual attention mechanism, and forces spatial relationship verification through an anatomical constraint loss function, thereby reducing the false detection rate caused by rib occlusion, eliminating anatomical structure interference, and reducing the false positive rate. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0029] Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0030] The present invention will be further described in detail below with reference to embodiments:

[0031] Example 1

[0032] like Figure 1 , Figure 2 As shown, this invention provides a PICC catheter recognition system based on image recognition, comprising: a medical image input module for acquiring X-ray image data; an improved YOLO-Seg segmentation module connected to the medical image input module, the improved YOLO-Seg segmentation module including: a CSPDarknet backbone network with the number of channels expanded to 1280, a feature pyramid network outputting 4-scale feature maps, and a segmentation head containing 32-channel prototype mask generation; a dual attention enhancement module integrated into the improved YOLO-Seg segmentation module, including a channel attention module and a spatial attention module, for enhancing the response of low-contrast catheter features; and an anatomical intelligent post-processing module for performing rib morphology optimization, catheter skeleton continuity reconstruction, and anatomical spatial relationship verification on the mask output by the segmentation module.

[0033] In this embodiment, the medical image input module uses a DICOM interface to receive X-ray images of 1024×1024px, and then uses bilinear interpolation to uniformly scale them down to 512×512px, preserving high-frequency details.

[0034] Example 2

[0035] like Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the loss function of the improved YOLO-Seg segmentation module includes: boundary detection loss, Focal Loss focusing on slender structures, boundary awareness loss, and anatomical loss based on rib position constraints. The anatomical intelligent post-processing module performs the following operations:

[0036] A1: Morphological optimization of the rib mask by sequentially performing opening and closing operations;

[0037] A2: Use skeleton analysis algorithm to connect broken segments on the PICC catheter mask;

[0038] A3: Correcting linear anatomical structures using Hough transform;

[0039] A4: Verify whether the spatial distance between the catheter and the rib is within the preset clinical safety threshold;

[0040] In the dual attention enhancement module: the channel attention module enhances the response of the catheter region by recalibrating the feature channel weights, and the spatial attention module generates a spatial weight map to focus on the anatomical edge.

[0041] In this embodiment, the backbone network of the improved YOLO-Seg segmentation module is: CSPDarknet with a depth multiplier of 1.0 and a width multiplier of 1.0, with the output channels expanded to 1280, and an embedded SPPF pyramid pooling layer. The segmentation head is a 32-channel prototype mask generator that receives the highest resolution feature map of 160×160×256 output by FPN. The channel attention in the dual attention enhancement module recalibrates the channel-level weights of the feature map output by the backbone network, with the formula $M_c(F)=\sigma(MLP(AvgPool(F))$). Spatial attention generates a spatial weight map, with the formula $M_s(F)=\sigma(Conv^{7×7}([AvgPool(F);MaxPool(F)]))$.

[0042] In the post-anatomical processing module, rib optimization: first, an opening operation is performed on the mask to eliminate noise, and then a closing operation is performed to fill the holes; duct reconstruction: the Zhang-Suen skeletonization algorithm is used to connect broken fragments and retain connected components with a length > 15 pixels.

[0043] Example 3

[0044] like Figure 1 , Figure 2 As shown, this invention provides a PICC catheter identification method based on image recognition, comprising the following steps:

[0045] S1: Acquire X-ray images and preprocess them to a resolution of 512×512;

[0046] S2: Object detection and instance segmentation are performed simultaneously by improving the YOLO-Seg model, where: a 32-channel prototype mask is used to generate slender duct segmentation results, and a dual attention mechanism is applied to enhance low-contrast region features;

[0047] S3: Intelligent post-processing of the output mask: Reconstructing duct continuity based on skeleton analysis and reducing false positives by utilizing rib spatial position constraints;

[0048] S4: Output the recognition results to the display terminal or clinical navigation system.

[0049] Example 4

[0050] like Figure 1 , Figure 2As shown, based on Example 3, this invention provides a technical solution: The improved YOLO-Seg model is optimized through multi-task collaborative training: the detection head and segmentation head are trained jointly, and an edge detection auxiliary task is added. The loss function includes an anatomical position constraint term, forcing the prediction results to conform to the clinical anatomical distribution. In the intelligent anatomical post-processing: duct continuity reconstruction uses pixel connectivity analysis and path search algorithms; spatial relationship verification rejects duct predictions located behind the ribs and without perspective overlap. It also includes a deployment optimization module, which performs FP16 quantization and layer fusion using TensorRT, supports dynamic input resolution of 512-1024px, and exports the ONNX format model to a medical edge device for real-time inference.

[0051] In this embodiment, the PyTorch model is converted to ONNX format, FP16 quantization is performed using the trtexec tool, and it is deployed on NVIDIA Jetson AGX Xavier. The measured inference speed is ≥38FPS.

[0052] The working principle of the PICC catheter identification method and system based on image recognition will be explained in detail below.

[0053] like Figure 1 , Figure 2 As shown, X-ray images are input, and after standardization at a resolution of 512×512px, pixel values ​​are normalized. Then, improved YOLO-Seg inference is used: multi-scale features are extracted from the CSPDarknet backbone and then FPN is used to fuse features. Then, the response of the catheter region is enhanced by a dual attention module. The prototype mask output by the 32-channel segmentation head is multiplied with the detection box matrix to generate an instance mask. Then, anatomical intelligent post-processing is performed: rib mask opening and closing operations smooth the edges, skeleton analysis ensures continuity, and false positive catheters with overlap with the ribs >15% or distance <5mm are rejected. The catheter centerline coordinates with confidence are sent to the ultrasound navigation system to guide clinical operation in real time.

[0054] The present invention has been described in detail above. However, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, any modifications or improvements that do not depart from the spirit of the present invention are within the scope of protection of the present invention.

Claims

1. An image recognition-based PICC catheter identification method, characterized in that: The method comprises the following steps: S1: Obtain X-ray images and preprocess to 512*512 resolution; S2: Perform target detection and instance segmentation simultaneously by improving the YOLO-Seg model, wherein: use a 32-channel prototype mask to generate an elongated catheter segmentation result, and apply a double attention mechanism to enhance the features of low-contrast areas; S3: Perform anatomical intelligent post-processing on the output mask: reconstruct the continuity of the catheter based on skeleton analysis, and reduce false positives by using rib spatial position constraints; S4: Output the recognition result to a display terminal or a clinical navigation system. 2.The PICC catheter identification method based on image recognition according to claim 1, characterized in that: The improved YOLO-Seg model is optimized through multi-task collaborative training: the detection head and the segmentation head are jointly trained, and an edge detection auxiliary task is added, and the loss function includes an anatomical position constraint term to force the prediction result to conform to the clinical anatomical distribution. 3.The PICC catheter identification method based on image recognition according to claim 1, characterized in that: In the anatomical intelligent post-processing: the catheter continuity reconstruction uses pixel connectivity analysis and path search algorithm, and the spatial relationship verification rejects the catheter prediction located behind the ribs and without perspective overlap.

4. An image recognition based PICC catheter identification system, characterized by: It comprises: a medical image input module for obtaining X-ray image data; an improved YOLO-Seg segmentation module connected to the medical image input module, the improved YOLO-Seg segmentation module comprising: a CSPDarknet backbone network with a channel number expanded to 1280, a feature pyramid network outputting 4 scale feature maps, and a segmentation head including a 32-channel prototype mask generation; a double attention enhancement module integrated in the improved YOLO-Seg segmentation module, comprising a channel attention module and a spatial attention module, for enhancing the low-contrast catheter feature response; an anatomical intelligent post-processing module for performing rib morphology optimization, catheter skeleton continuity reconstruction, and anatomical spatial relationship verification on the mask output by the segmentation module.

5. The image recognition based PICC catheter identification system of claim 4, wherein: The loss function of the improved YOLO-Seg segmentation module includes: boundary detection loss, Focal Loss focusing on elongated structures, boundary perception loss, and anatomical loss based on rib position constraints.

6. The image recognition based PICC catheter identification system of claim 4, wherein: The anatomical intelligent post-processing module performs the following operations: A1: Perform morphological optimization of the rib mask by sequentially performing opening operation and closing operation; A2: Use skeleton analysis algorithm to connect the broken fragments of the PICC catheter mask; A3: Correct the straight-line anatomical structure by Hough transform; A4: Check whether the spatial distance between the catheter and the rib is within the preset clinical safety threshold.

7. The image recognition based PICC catheter identification system of claim 4, wherein: In the double attention enhancement module: the channel attention module enhances the catheter area response by recalibrating the feature channel weight, and the spatial attention module generates a spatial weight map to focus on the anatomical edge.

8. The image recognition based PICC catheter identification system of claim 4, wherein: It also comprises a deployment optimization module, which performs FP16 quantization and layer fusion through TensorRT, supports dynamic input resolution of 512-1024px, and exports the ONNX format model to medical edge devices for real-time inference.

Images