Placenta defect detection method and device based on multispectral imaging, equipment and medium

Abstract

The present application relates to the technical field of medical image processing, and particularly relates to a placenta defect detection method and device based on multispectral imaging, equipment and medium, the method first acquires the RGB image and infrared image of the placenta, generates a target fusion image based on the RGB image and infrared image, then calls a pre-trained semantic segmentation model to perform semantic segmentation processing on the target fusion image, obtains a placenta maternal surface area image, performs defect recognition processing on the placenta maternal surface area image, obtains a binary defect mask, constructs a height map based on the infrared image, and finally determines the defect area of the placenta based on the binary defect mask and the height map, and performs visual processing on the defect area to obtain a detection report, aiming to realize the objectification and automation of placenta integrity inspection, and effectively solve the problems of low reliability, poor efficiency, high risk and the like caused by relying on manual operation in the existing placenta integrity inspection.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method, apparatus, equipment and medium for detecting placental defects based on multispectral imaging. Background Technology

[0002] The placenta is the core organ for the exchange of substances between the mother and fetus during pregnancy. It must be delivered intact during vaginal delivery or cesarean section. Residual placental tissue in the uterine cavity is a major cause of late postpartum hemorrhage, with an overall incidence of approximately 0.5%–3%, which is significantly higher in high-risk women with placenta previa or a history of multiple curettages. Therefore, relevant obstetric and gynecological monographs and international obstetric guidelines clearly state that the integrity of the placenta should be checked immediately after delivery, with a focus on the maternal side facing the uterine wall. The normal maternal side consists of 15–20 placental lobules and appears as dark red, rough, granular particles. The absence of lobules will appear as a smooth, concave area, indicating that the area may still be attached to the uterine wall.

[0003] Currently, the examination of placental integrity worldwide still relies entirely on manual operation. After the midwife lays the placenta flat, she judges whether there are any defects by touch and visual inspection. This method does not require instruments and is inexpensive, but it has many fundamental limitations. The results of manual examination are greatly affected by the operator's experience, have low sensitivity in identifying small defects, and cannot quantify the condition of defects. At the same time, there are no objective images and data records during the examination process. Furthermore, in high-pressure clinical scenarios such as precipitous labor and multiple deliveries, the examination procedure is easily simplified, and the judgment criteria of different personnel lack uniformity. If postpartum hemorrhage occurs, it is impossible to trace the true state of the placenta at the time of delivery. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method, device, equipment and medium for placental defect detection based on multispectral imaging, which aims to realize the objectivity and automation of placental integrity examination, and effectively solve the problems of low reliability, poor efficiency and high risk caused by reliance on manual operation in the existing placental integrity examination.

[0005] The first aspect of this invention provides a method for detecting placental defects based on multispectral imaging, comprising: acquiring RGB and infrared images of the placenta; generating a target fusion image based on the RGB and infrared images; performing semantic segmentation processing on the target fusion image using a pre-trained semantic segmentation model to obtain a placental maternal surface region image; performing defect identification processing on the placental maternal surface region image to obtain a binary defect mask; constructing a height map based on the infrared image; determining the placental defect region based on the binary defect mask and the height map, and visualizing the defect region to obtain a detection report.

[0006] Optionally, in a first implementation of the first aspect of the present invention, generating a target fused image based on the RGB image and the infrared image includes: performing spatial alignment processing on the RGB image and the infrared image using an image registration algorithm to obtain a dual-channel aligned image; performing feature fusion processing on the dual-channel aligned image using a channel stitching fusion algorithm to obtain an initial fused image; performing noise reduction processing on the initial fused image using a filtering algorithm to obtain a noise-reduced fused image; and performing contrast enhancement processing on the noise-reduced fused image using a histogram equalization algorithm to obtain the target fused image.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the semantic segmentation model includes an encoder module, a decoder module, and a mask generation module, wherein the encoder module, the decoder module, and the mask generation module are connected in sequence. The step of calling the pre-trained semantic segmentation model to perform semantic segmentation processing on the target fused image to obtain a placental maternal surface region image includes: performing multi-scale downsampling feature extraction processing on the target fused image based on the encoder module to obtain a compressed semantic feature map; performing feature fusion processing on the compressed semantic feature map based on the decoder module to obtain a high-resolution feature map; and performing binarization region extraction processing on the high-resolution feature map based on the mask generation module to obtain the placental maternal surface region image.

[0008] Optionally, in a third implementation of the first aspect of the present invention, the step of performing defect identification processing on the placental maternal surface region image to obtain a binary defect mask includes: acquiring all pixels of the placental maternal surface region image, a predefined mean calculation function, a predefined standard deviation calculation function, a preset grayscale mean threshold, and a preset texture standard deviation threshold; acquiring multiple neighborhood windows of preset sizes centered on each pixel; calculating the local grayscale mean of each neighborhood window based on the predefined mean calculation function; calculating the local grayscale standard deviation of each neighborhood window based on the predefined standard deviation calculation function; comparing each local grayscale mean with the grayscale mean threshold, and comparing each local grayscale standard deviation with the texture standard deviation threshold; determining pixels whose local grayscale mean is less than the grayscale mean threshold and whose local grayscale standard deviation is less than the texture standard deviation threshold as defect points, and generating the binary defect mask based on the defect points.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, generating the binary defect mask based on the defect points includes: generating a binary image based on the defect points using a pixel-level assignment algorithm; performing an opening operation on the binary image using a morphological opening operation algorithm to obtain a cleaned binary image; performing a closing operation on the cleaned binary image using a morphological closing operation algorithm to obtain a smoothed binary image; and performing connected component analysis on the smoothed binary image using a connected component filtering algorithm to obtain the binary defect mask.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, the step of constructing a height map based on the infrared image includes: obtaining a preset reflectance, a predefined mapping function, and a predefined tray plane height reference parameter; constructing a Lambertian reflection model based on the reflectance and the infrared image, and calculating the placental surface normal vector based on the Lambertian reflection model; calculating the placental surface gradient field using the mapping function based on the placental surface normal vector; performing an integral algorithm to integrate the placental surface gradient field to obtain a placental surface height map; and performing global calibration processing on the placental surface height map based on the tray plane height reference parameter to obtain the height map.

[0011] Optionally, in a sixth implementation of the first aspect of the present invention, determining the defect area of ​​the placenta based on the binary defect mask and the height map includes: obtaining preset tray center coordinates, a standard height average, and a defect height value corresponding to each defect point in the binary defect mask; defining the difference between each defect height value and the standard height average as the depression depth of the corresponding defect point, obtaining multiple depression depths, and determining a maximum depression depth and an average depression depth based on the multiple depression depths; calculating the geometric centroid coordinates based on the binary defect mask, and determining spatial location information based on the distance and angle between the geometric centroid coordinates and the tray center coordinates; and determining the defect area based on the maximum depression depth, the average depression depth, and the spatial location information.

[0012] A second aspect of the present invention provides a placental defect detection device based on multispectral imaging, comprising: an image fusion module for acquiring RGB and infrared images of the placenta, and generating a target fused image based on the RGB and infrared images; a semantic segmentation module for calling a pre-trained semantic segmentation model to perform semantic segmentation processing on the target fused image to obtain a placental maternal surface region image; a defect identification module for performing defect identification processing on the placental maternal surface region image to obtain a binary defect mask; a height map construction module for constructing a height map based on the infrared image; and a defect region determination module for determining the placental defect region based on the binary defect mask and the height map, and visualizing the defect region to obtain a detection report.

[0013] A third aspect of the present invention provides a placental defect detection device based on multispectral imaging, the placental defect detection device based on multispectral imaging comprising: a memory and at least one processor, the memory storing instructions; at least one processor calling the instructions in the memory to cause the placental defect detection device based on multispectral imaging to perform the various steps of the placental defect detection method based on multispectral imaging described above.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the placental defect detection method based on multispectral imaging described in any of the preceding claims.

[0015] In the technical solution of this invention, firstly, RGB and infrared images of the placenta are acquired. A target fusion image is generated based on the RGB and infrared images. Then, a pre-trained semantic segmentation model is called to perform semantic segmentation processing on the target fusion image to obtain a placental maternal surface region image. Defect recognition processing is performed on the placental maternal surface region image to obtain a binary defect mask. A height map is constructed based on the infrared image. Finally, the defect area of ​​the placenta is determined based on the binary defect mask and the height map, and the defect area is visualized to obtain a detection report. This aims to achieve objectivity and automation of placental integrity inspection, effectively solving the problems of low reliability, poor efficiency, and high risk caused by reliance on manual operation in existing placental integrity inspections. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the placental defect detection method based on multispectral imaging provided in this embodiment of the invention; Figure 2 A schematic diagram of the structure of the placental defect detection device based on multispectral imaging provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a placental defect detection device based on multispectral imaging provided in an embodiment of the present invention. Detailed Implementation

[0017] This invention provides a method, apparatus, device, and medium for detecting placental defects based on multispectral imaging. In this invention, the terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the placental defect detection method based on multispectral imaging in this invention includes: 101. Acquire RGB and infrared images of the placenta, and generate a target fusion image based on the RGB and infrared images; In this embodiment, the acquisition of RGB and infrared images of the placenta is achieved using a dedicated tray module and a multispectral imaging module. The dedicated tray module is made of white matte medical plastic, measuring 25cm × 25cm, with a smooth, non-reflective surface. The multispectral imaging module is embedded in the bottom, with the lens facing upwards, covering the entire tray area. It is detachable and disposable. The multispectral imaging module includes two light source channels and a high-resolution sensor. The visible light channel uses a 400-700nm white LED to acquire the color and texture of the placental surface. The near-infrared channel uses an 850nm LED to enhance tissue blood supply differences and highlight the texture contrast between normal placental lobules and defective areas. The sensor resolution is 1920 × 1080, with a frame rate of no less than 15fps, and the dual-channel images are transmitted synchronously via a USB interface.

[0019] In this embodiment, after the two types of images are acquired synchronously, an image registration algorithm is used to complete spatial alignment processing to form a dual-channel aligned image, ensuring that the pixel coordinates of the two types of images correspond accurately. On this basis, a channel stitching and fusion algorithm is used to perform feature fusion on the dual-channel aligned image, integrating complementary information from the visible light and near-infrared bands to obtain an initial fused image. Then, a filtering algorithm is used to perform noise reduction processing on the initial fused image to eliminate environmental noise and stray light interference during the imaging process, resulting in a noise-reduced fused image. Finally, a histogram equalization algorithm is used to perform contrast enhancement processing on the noise-reduced fused image, strengthening the grayscale differences between placental tissue and background, and between normal tissue and defective areas, ultimately obtaining the target fused image.

[0020] 102. Call the pre-trained semantic segmentation model to perform semantic segmentation processing on the target fused image to obtain the placental maternal surface region image; In this embodiment, a pre-trained semantic segmentation model is invoked to perform semantic segmentation processing on the target fused image. This model is a lightweight U-Net structure, which includes an encoder module, a decoder module, and a mask generation module connected in sequence. During processing, the encoder module first performs multi-scale downsampling feature extraction on the target fused image, compressing the image spatial dimension while preserving the core semantic information of the placental tissue, and obtaining a compressed semantic feature map. Then, the decoder module performs feature fusion processing on the compressed semantic feature map, and restores the image resolution through upsampling decoding and feature mapping fusion to obtain a high-resolution feature map. Finally, the mask generation module performs binarization region extraction on the high-resolution feature map and outputs a placental maternal surface region mask. During the process, the texture differences of placental tissue under the near-infrared band can be highlighted, effectively eliminating interference from non-target areas such as the umbilical cord and amnion, and finally obtaining a clean placental maternal surface region image. This semantic segmentation process relies on the encoding and decoding structure of the lightweight U-Net model, which can efficiently extract multi-scale semantic features and accurately restore the spatial contour and details of the placental maternal surface. By extracting binary regions, it can achieve accurate separation between target regions and interference regions, providing high-quality and clean input for subsequent placental defect identification. This can significantly improve the accuracy and robustness of defect detection and avoid interference from non-target regions on defect identification results.

[0021] 103. Perform defect recognition processing on the maternal surface region image of the placenta to obtain a binary defect mask; In this embodiment, a neighborhood window of a preset size is selected, centered on each pixel within the maternal surface region image of the placenta. The size of this neighborhood window is adapted to the typical structural features of the placental lobules, accurately capturing the local grayscale and texture information of individual lobules and defective areas. Based on a predefined mean calculation function, the pixel grayscale values ​​within each neighborhood window are statistically calculated to obtain the local grayscale mean of the corresponding pixel. Simultaneously, relying on a predefined standard deviation calculation function, the deviation between the pixel grayscale values ​​within each neighborhood window and the local grayscale mean is calculated, yielding the local grayscale standard deviation reflecting texture roughness. This calculation process quantifies the local features of the placental tissue, providing quantitative data support for subsequent discrimination.

[0022] After completing the local feature calculation, the local gray-level mean of each pixel is compared with the preset gray-level mean threshold, and the local gray-level standard deviation is compared with the preset texture standard deviation threshold. When a pixel meets the conditions that the local gray-level mean is less than the gray-level mean threshold and the local gray-level standard deviation is less than the texture standard deviation threshold, the pixel is determined to be a defect. The thresholds based on clinical samples can effectively distinguish between normal tissue and defective tissue, improving the accuracy and objectivity of defect identification. When generating a binary defect mask based on the identified defect points, a pixel-level assignment algorithm is first used to generate the corresponding binary image. Then, a morphological opening algorithm is used to process the binary image, removing isolated noise points to purify the image. Subsequently, a morphological closing algorithm is used to process the purified binary image, filling in small internal holes to optimize the image morphology. Finally, a connected component filtering algorithm is used to perform connected component analysis, eliminating tiny connected components with areas smaller than a preset threshold to prevent non-defective micro-regions such as bloodstains and bubbles from being misjudged as defects. The final result is a well-formed and accurately characterized binary defect mask. The regions with a value of 1 in the mask represent the placental tissue defect areas identified by the system. Compared to traditional manual examinations that rely on subjective experience, this method achieves automated and standardized identification of placental maternal surface defect areas. It not only improves the sensitivity for identifying small defects but also provides objective binary mask representations, laying a high-quality foundation for subsequent three-dimensional quantitative analysis and visualization of defect areas. This effectively solves the technical shortcomings of existing manual inspections, which lack quantification and objective recording, significantly improving the overall accuracy and reliability of placental integrity detection.

[0023] 104. Construct a height map based on the infrared image; In this embodiment, a Lambertian reflection model is first constructed based on the acquired preset reflectivity and the brightness information of the near-infrared image. This model uses the vertically upward parallel light source condition guaranteed by the device structure as the light source condition, and the placental surface conforms to the diffuse reflection physical characteristics of a Lambertian body. The Lambertian reflection model is used to calculate the placental surface normal vector. The placental surface normal vector serves as a geometric quantity characterizing the three-dimensional orientation of the placental surface, providing crucial physical and geometric basis for the subsequent gradient field solution. Subsequently, a predefined mapping function is used to calculate the placental surface gradient field based on the calculated placental surface normal vector. This mapping function establishes a quantitative correlation between the normal vector and the surface gradient, converting the normal vector information into gradient field data that can be integrated, thus achieving an effective conversion from the two-dimensional brightness information of the near-infrared image to the three-dimensional gradient information of the placental surface. After calculating the gradient field, an integral algorithm is used to integrate the gradient field on the placental surface. This algorithm, based on the Frankot-Chellappa algorithm, integrates the gradient field information to restore the height information of the placental surface, resulting in a placental surface height map. Finally, the placental surface height map is globally calibrated according to a predefined tray plane height reference parameter. This parameter uses the tray plane as the zero point of height, eliminating accumulated errors during integration and giving the height map a true physical scale meaning. Ultimately, a height map representing the three-dimensional morphology of the placental surface is obtained. Compared to traditional manual inspection methods that rely solely on two-dimensional morphological observation, this method enables the digital reconstruction of the three-dimensional morphology of the placental surface. It reflects the concavity and convexity features of the placental surface and the three-dimensional morphological differences of defect areas, providing high-precision three-dimensional data support for subsequent quantitative analysis of the size and depth of defect areas. This effectively improves the quantitative level and accuracy of placental defect detection and fills the gap in three-dimensional morphological analysis in placental integrity detection technology, providing objective and accurate evidence for placental morphology detection in obstetric clinics.

[0024] 105. Based on the binary defect mask and the height map, determine the defect area of ​​the placenta, and perform visualization processing on the defect area to obtain a detection report.

[0025] In this embodiment, the preset tray center coordinates and the defect height value corresponding to each defect point in the binary defect mask are first obtained. For each defect point, within a defined annular neighborhood, normal tissue pixels located within the maternal surface mask are searched. The average height of the normal tissue in this neighborhood is calculated as the standard height mean. The difference between the defect height value of each defect point and the corresponding standard height mean is defined as the depression depth of that point. After traversing all defect points to obtain all depression depths, the maximum value is taken as the maximum depression depth, and the average value is taken as the average depression depth, thus completing the quantitative characterization of the depression degree of the defect area. Subsequently, the geometric centroid coordinates of the defect area are calculated based on the binary defect mask. Using the preset tray center coordinates as the origin, the distance and angle between the centroid and the tray center are calculated. The angle is converted into a clock orientation form that conforms to clinical description habits to obtain the spatial location information of the defect area. Combining the maximum depression depth, the average depression depth, and the spatial location information, the complete determination of the placental defect area is completed. In the visualization process, a pseudo-color heatmap is overlaid on the original RGB image, highlighting the defect area in red. The deeper the depression, the darker the corresponding color, achieving an intuitive visualization of the defect area. If the defect area or depth exceeds a preset threshold, the system will automatically pop up a warning message, prompting the clinician to conduct a uterine cavity exploration. The user can trigger a save operation, and the system will automatically generate a PDF test report containing the image, quantitative parameters, and timestamps. The preset threshold is set by collecting a large number of clinical samples of placental defects with different gestational weeks, different delivery methods, and different high-risk factors. The correlation between the defect area, depression depth, and the risk of complications such as late postpartum hemorrhage and retained products of conception is statistically analyzed to determine clinically significant critical values, which serve as the basic data support for threshold setting. At the same time, the structured data is uploaded to the hospital's EMR system through the communication module to complete the retention and archiving of the test data. It enables multi-dimensional quantitative analysis and intuitive visualization of placental defect areas. Compared with traditional manual examination, it can accurately quantify the depth and spatial location of the defect, avoiding errors caused by subjective judgment. At the same time, the generated test report and uploaded structured data provide objective basis for clinical decision-making, and facilitate medical quality control and dispute evidence collection, effectively improving the standardization and intelligence level of placental integrity testing.

[0026] In this embodiment of the invention, generating a target fused image based on the RGB image and the infrared image includes: using an image registration algorithm to spatially align the RGB image and the infrared image to obtain a dual-channel aligned image; using a channel stitching fusion algorithm to perform feature fusion on the dual-channel aligned image to obtain an initial fused image; using a filtering algorithm to denoise the initial fused image to obtain a denoised fused image; and using a histogram equalization algorithm to enhance the contrast of the denoised fused image to obtain the target fused image.

[0027] In this embodiment, the RGB and infrared images acquired by the multispectral imaging module exhibit spatial offset due to light source switching or hardware position differences. An image registration algorithm is used to spatially align the two types of images, resulting in a dual-channel aligned image. Specifically, SIFT or ORB algorithms based on feature points can be employed. By extracting stable edge or texture feature points from both types of images, matching corresponding features, and calculating the transformation matrix, pixel-level coordinate alignment is achieved, ensuring precise spatial correspondence between the two types of images and providing a reliable foundation for subsequent fusion. The dual-channel aligned image is then subjected to feature fusion using a channel stitching fusion algorithm to obtain an initial fused image. Specifically, a channel cascading fusion method can be used, directly stitching the three channels of the RGB image with the single channel of the infrared image into a four-channel feature map. This integrates the color and texture information of visible light with the tissue blood supply difference information of near-infrared light, highlighting the texture contrast between normal placental lobules and defective areas in the near-infrared band, providing richer feature dimensions for subsequent processing. The initial fused image is denoised using a filtering algorithm to obtain a denoised fused image. Gaussian filtering or median filtering can be used. Gaussian filtering smooths Gaussian noise while preserving the overall image texture, while median filtering effectively removes salt-and-pepper noise and eliminates noise generated by ambient light and electronic interference during imaging, thus improving image quality and preventing noise interference with subsequent feature extraction and segmentation. The denoised fused image is then contrast-enhanced using a histogram equalization algorithm. This involves adjusting the grayscale distribution in local areas to strengthen the grayscale differences between placental tissue and background, and between normal tissue and defective areas. This further highlights the texture differences in the near-infrared region, improving the image's visual recognizability and feature separability. The resulting target fused image retains the morphological details of the placental surface while highlighting the texture contrast caused by differences in tissue blood supply. This provides clear, high-quality input for subsequent placental maternal surface region extraction and defect identification, effectively improving the accuracy and robustness of subsequent processing. It solves the problem of insufficient information in single-band images, and the standardized algorithm process ensures the stability and repeatability of the processing.

[0028] In this embodiment of the invention, the semantic segmentation model includes an encoder module, a decoder module, and a mask generation module, which are sequentially connected. The step of calling the pre-trained semantic segmentation model to perform semantic segmentation processing on the target fused image to obtain a placental maternal surface region image includes: performing multi-scale downsampling feature extraction processing on the target fused image based on the encoder module to obtain a compressed semantic feature map; performing feature fusion processing on the compressed semantic feature map based on the decoder module to obtain a high-resolution feature map; and performing binarization region extraction processing on the high-resolution feature map based on the mask generation module to obtain the placental maternal surface region image.

[0029] In this embodiment, a pre-trained semantic segmentation model is used to perform semantic segmentation processing on the target fused image to obtain the placental maternal surface region image. The model is a lightweight U-Net structure, which includes an encoder module, a decoder module and a mask generation module connected in sequence. The overall processing flow takes the target fused image as input and sequentially completes multi-scale feature extraction, feature fusion and binarization region extraction, and finally achieves accurate separation of the placental maternal surface region and non-target regions. The semantic segmentation model is trained using a placental multispectral image dataset. First, a training dataset containing numerous fused placental images and corresponding pixel-level labeled masks is constructed. These labeled masks must clearly distinguish the placental maternal surface region from non-target regions such as the umbilical cord and amnion, ensuring accuracy and completeness to support the model's feature learning. Then, the training dataset is input into a lightweight U-Net model for training. An adaptive moment estimation algorithm is used as the optimizer, combined with a loss function that integrates cross-entropy and Dice loss functions. This combined loss function effectively addresses the pixel imbalance between the placental maternal surface region and the background region, improving the model's segmentation ability for small-area maternal surface regions. During training, data augmentation strategies are implemented simultaneously, such as random flipping, rotation, and brightness adjustment, to expand the dataset sample size and improve the model's generalization ability. The model's segmentation performance is monitored in real-time using a validation set. Training stops when the model's segmentation metrics on the validation set reach a preset threshold, resulting in a stable pre-trained semantic segmentation model that can be directly applied to semantic segmentation of the placental maternal surface region.

[0030] In this embodiment, the encoder module, as the core unit for feature extraction, performs multi-scale downsampling feature extraction on the target fused image. Through continuous convolution and downsampling operations, it progressively compresses the spatial dimension of the image while capturing the semantic features of placental tissue at different scales. This includes the overall structural features of the placental maternal surface, lobular texture features, and edge features of non-target areas such as the umbilical cord and amnion. The final output is a compressed semantic feature map that retains core semantic information while compressing dimensions, providing a representative feature foundation for subsequent processing. The decoder module performs feature fusion processing on the compressed semantic feature map output by the encoder module. Through upsampling operations, it progressively restores the spatial resolution of the image. Simultaneously, it combines the feature information from the encoder module at the corresponding scale to perform cross-layer feature fusion, effectively integrating shallow detail features with deep semantic features. This compensates for spatial details lost during downsampling, ultimately resulting in a high-resolution feature map with rich semantic information and clear spatial details, capable of accurately reconstructing the contour and internal structural features of the placental maternal surface. The mask generation module performs binarization and region extraction processing on the high-resolution feature map output by the decoder module. Through pixel-level classification, it divides the pixels in the feature map into the placental maternal surface region and non-target regions, outputting the corresponding placental maternal surface region mask. This achieves accurate separation of the target region and interference regions, ultimately yielding a clean image of the placental maternal surface region, completing the entire semantic segmentation process. This provides a clean and reliable input foundation for subsequent placental defect identification, 3D depth estimation, and quantitative analysis of defect regions, effectively improving the overall accuracy and robustness of placental integrity detection. In this embodiment of the invention, the defect identification processing of the placental maternal surface region image to obtain a binary defect mask includes: acquiring all pixels of the placental maternal surface region image, a predefined mean calculation function, a predefined standard deviation calculation function, a preset grayscale mean threshold, and a preset texture standard deviation threshold; acquiring multiple neighborhood windows of preset sizes centered on each pixel; calculating the local grayscale mean of each neighborhood window based on the predefined mean calculation function; calculating the local grayscale standard deviation of each neighborhood window based on the predefined standard deviation calculation function; comparing each local grayscale mean with the grayscale mean threshold, and comparing each local grayscale standard deviation with the texture standard deviation threshold; determining pixels whose local grayscale mean is less than the grayscale mean threshold and whose local grayscale standard deviation is less than the texture standard deviation threshold as defect points, and generating the binary defect mask based on the defect points.

[0031] In this embodiment, defect identification processing is performed on the maternal surface region image of the placenta to obtain a binary defect mask. First, all pixels of the maternal surface region image of the placenta, a predefined mean calculation function, a predefined standard deviation calculation function, a preset gray-level mean threshold, and a preset texture standard deviation threshold need to be obtained. The predefined mean calculation function is used to quantify the average gray level of the local area, and the predefined standard deviation calculation function is used to reflect the texture roughness of the local area. The preset gray-level mean threshold and texture standard deviation threshold are both obtained based on clinical sample calibration, which can effectively distinguish between normal placental lobule regions and defective regions.

[0032] In this embodiment, a neighborhood window of a preset size is obtained centered on each pixel. This window size is adapted to the typical structure of a placental lobule, typically using a 7×7 specification, which can accurately capture the local grayscale and texture features of a single lobule and its surrounding area. The local grayscale mean of each neighborhood window is calculated based on a predefined mean calculation function, and the corresponding formula expression is: , in, Represented by pixels The local gray-scale mean value centered on the center, 49 represents The total number of pixels within the neighborhood window Represented by pixels Centered Neighborhood window set, Represents pixels within the neighborhood window The corresponding near-infrared image grayscale value can be used to calculate the average grayscale level of the local area, providing basic data for subsequent discrimination.

[0033] The local grayscale standard deviation of each neighborhood window is calculated based on a predefined standard deviation calculation function, and the corresponding formula expression is: , in, Represented by pixels The local grayscale standard deviation centered on the center is used to reflect the texture roughness of a local area; 48 represents... The degrees of freedom of the neighborhood window (obtained by subtracting 1 from the total number of window pixels, 49). Represented by pixels The local grayscale mean value centered on Represents pixels within the neighborhood window The corresponding near-infrared image grayscale value, Represented by pixels Centered Neighborhood window set; the smaller the local grayscale standard deviation, the smoother the texture of the local area, and vice versa.

[0034] In this embodiment, the mean gray level of each local area is compared with a preset mean gray level threshold, and the standard deviation of each local gray level is compared with a preset texture standard deviation threshold. The preset mean gray level threshold is usually set to 80 (to adapt to the gray level range of 0-255 in an 8-bit image), and the preset texture standard deviation threshold is usually set to 15. When a pixel meets the conditions that the mean gray level of a local area is less than the mean gray level threshold and the standard deviation of a local gray level is less than the texture standard deviation threshold, the pixel is determined to be a defect. This discrimination logic is consistent with the near-infrared image characteristics of normal placental lobule regions having high gray level values ​​and rough texture, and defective regions having low gray level values ​​and smooth surfaces, which can accurately identify suspected defective regions. Based on the identified defect points, a temporary binary image is generated, with defect points marked as 1 and non-defect points marked as 0. Then, morphological operations are used to clean the temporary binary image. First, opening operations are used to remove isolated noise points, then closing operations are used to fill small holes inside, and finally, connected component analysis is performed to remove tiny regions with too small an area, so as to avoid non-defective tiny regions such as bloodstains and bubbles being misjudged as defects. Finally, a binary defect mask with regular shape and accurate features is obtained. The regions with a value of 1 in the mask are the placental tissue defect regions identified by the system.

[0035] In this embodiment of the invention, generating the binary defect mask based on the defect points includes: generating a binary image based on the defect points using a pixel-level assignment algorithm; performing an opening operation on the binary image using a morphological opening operation algorithm to obtain a cleaned binary image; performing a closing operation on the cleaned binary image using a morphological closing operation algorithm to obtain a smoothed binary image; and performing connected component analysis on the smoothed binary image using a connected component filtering algorithm to obtain the binary defect mask.

[0036] In this embodiment, a binary defect mask is generated based on the defect points. First, a pixel-level assignment algorithm is used to generate a binary image based on the defect points. This algorithm assigns a value of 1 to the identified defect point pixels and a value of 0 to the non-defect point pixels, generating a temporary binary image to achieve the binary representation of the defect points. Then, a morphological opening operation algorithm is used to perform opening operations on the binary image. This algorithm follows an operation flow of erosion followed by dilation, using 3×3 circular structuring elements for the operation. The erosion operation shrinks the highlighted areas in the binary image, effectively removing isolated noise points not connected to the defect area. The dilation operation restores the original size of the defect area, ultimately obtaining a cleaned binary image that eliminates the interference of small noise points on subsequent processing. Next, a morphological closing operation algorithm is used to perform closing operations on the cleaned binary image. This algorithm follows an operation flow of dilation followed by erosion. The dilation operation expands the highlighted range of the defect area, filling in the small holes inside the area. The erosion operation shrinks the highlighted area back to its original size, resulting in a smoothed binary image that makes the outline of the defect area more continuous and complete. Finally, a connected component filtering algorithm is used to perform connected component analysis on the smoothed binary image. This algorithm first marks all independent highlighted connected regions using the 8-connectivity rule, calculates the pixel area of ​​each connected region, and then removes tiny regions with areas smaller than a preset threshold (which can be set to 25 pixels, approximately 0.25 cm²) to avoid misjudging non-defective tiny areas such as bloodstains and bubbles as defects. The final result is a binary defect mask, where the regions with a value of 1 are the placental tissue defect areas identified by the system. Pixel-level assignment achieves digital representation of the defect points. The combination of morphological opening and closing operations not only purifies image noise but also optimizes the morphological integrity of the defect area. Connected component filtering further eliminates clinically meaningless tiny interference regions. The resulting binary defect mask has high accuracy and robustness, accurately reflecting the true shape and extent of placental defects.

[0037] In this embodiment of the invention, constructing a height map based on the infrared image includes: obtaining a preset reflectance, a predefined mapping function, and a predefined tray plane height reference parameter; constructing a Lambertian reflection model based on the reflectance and the infrared image, and calculating the placental surface normal vector based on the Lambertian reflection model; calculating the placental surface gradient field using the mapping function based on the placental surface normal vector; performing an integral algorithm to integrate the placental surface gradient field to obtain a placental surface height map; and performing global calibration processing on the placental surface height map according to the tray plane height reference parameter to obtain the height map.

[0038] In this embodiment, to construct a height map based on infrared images, it is necessary to first obtain a preset reflectance, a predefined mapping function, and a predefined tray plane height reference parameter. The preset reflectance is the diffuse reflectance of placental tissue, which is obtained by pre-calibration of clinical samples. The predefined mapping function is the quantitative conversion relationship between the placental surface normal vector and the surface gradient field. The predefined tray plane height reference parameter is the height value of the tray plane, which is set to 0 mm as the physical zero point for height calculation.

[0039] In this embodiment, a Lambertian reflection model is constructed based on preset reflectance and infrared images. First, the core premise and preset parameters for model construction are clarified. The preset reflectance is the diffuse reflectance of placental tissue in the near-infrared band, obtained through extensive experimental calibration of the near-infrared reflectance characteristics of clinical placental samples. This preset parameter is used to characterize the reflective optical properties of the placental tissue itself, ensuring that the model closely matches the actual diffuse reflectance characteristics of the tissue. Simultaneously, the device structure ensures that the light source is vertically upward parallel light, corresponding to the light source direction vector. The value is (0,0,1), and this vector is a preset geometric parameter at the hardware level to ensure the determinism of the light source direction. Furthermore, based on the optical properties of placental tissue, it is assumed that the placental surface is a Lambertian material, satisfying the physical characteristics of diffuse reflection. Its surface brightness is only related to the angle between the light source direction and the surface normal vector; this is the physical assumption for constructing the Lambertian reflection model. Based on this, a model is constructed by combining pixel brightness information obtained from near-infrared image acquisition. The near-infrared image brightness... Represents the near-infrared image in pixel coordinates The brightness value at that location is a measured input parameter, representing the near-infrared light reflection intensity at that position. Based on Lambert's law of reflection, combined with a preset reflectance... Light source direction vector and the placental surface normal vector to be solved Construct a quantitative expression for the Lambert reflection model, the corresponding formula is: Among them, the placental surface normal vector is composed of , and It consists of three components.

[0040] In this embodiment, a predefined mapping function is used to calculate the placental surface gradient field based on the placental surface normal vector. The corresponding formula for this mapping function is: ,in, Indicates the placental surface is Gradient of direction, Indicates the placental surface is gradient of direction, and Representing the height field respectively exist and The partial derivatives of the direction can be used through this mapping function to convert three-dimensional normal vector information into integrable two-dimensional gradient field data, achieving a precise conversion from normal vector to gradient field.

[0041] In this embodiment, an integral algorithm is used to calculate the gradient field on the placental surface. Specifically, the Frankot-Chellappa algorithm is used to calculate the gradient field. and By performing two-dimensional integration, the height field of the continuous placental surface is obtained. The initial placental surface height map, measured in millimeters, is then discretely sampled into pixel-level height data. This algorithm effectively handles the integral constraint of the gradient field, ensuring the stability and accuracy of the height reconstruction. The initial placental surface height map is then globally calibrated based on predefined tray plane height reference parameters. These parameters are set to a height of 0 millimeters at the tray plane. By calibrating the height values ​​of the corresponding areas of the tray plane in the height map to 0, accumulated errors generated during integration are eliminated, giving the height map values ​​true physical scale significance. Finally, a height map accurately representing the three-dimensional morphology of the placental surface is obtained.

[0042] In this embodiment, the construction process fully relies on the physical principle of Lambertian reflection, combining the preset reflectivity, the hardware-preset light source direction, and the measured infrared image brightness to construct a Lambertian reflection model that conforms to the optical characteristics of placental tissue. This avoids the arbitrariness of model parameters and ensures the practical applicability of the model through measured image data. It achieves a precise correlation between the two-dimensional brightness information of the infrared image and the three-dimensional geometric features of the placental surface, providing a rigorous theoretical and computational basis for subsequent gradient field calculation and height map reconstruction based on normal vectors. This effectively ensures the accuracy and physical rationality of the three-dimensional morphology reconstruction of the placental surface and solves the technical problem that relying solely on image brightness cannot characterize the three-dimensional structure of the surface.

[0043] In this embodiment of the invention, determining the defect area of ​​the placenta based on the binary defect mask and the height map includes: obtaining preset tray center coordinates, standard height average, and defect height value corresponding to each defect point in the binary defect mask; defining the difference between each defect height value and the standard height average as the depression depth of the corresponding defect point, obtaining multiple depression depths, and determining the maximum depression depth and average depression depth based on the multiple depression depths; calculating the geometric centroid coordinates based on the binary defect mask, and determining spatial location information based on the distance and angle between the geometric centroid coordinates and the tray center coordinates; and determining the defect area based on the maximum depression depth, average depression depth, and spatial location information.

[0044] In this embodiment, determining the placental defect area based on a binary defect mask and height map requires first obtaining preset tray center coordinates, standard average height, and the defect height value corresponding to each defect point in the binary defect mask. The preset tray center coordinates are the geometric center coordinates of the tray plane, which are fixed parameters obtained through hardware calibration before the device leaves the factory, and serve as the origin for spatial location calculation. The standard average height is not a preset fixed value, but rather, for each defect point, the average height of normal tissue pixels located within the maternal surface area is calculated by filtering out pixels within a preset radius annular neighborhood around it.

[0045] In this embodiment, the defect height is the value of each pixel marked as defective in the binary defect mask. The corresponding height value in the height map, i.e. For each missing point, calculate its corresponding average standard height, as shown in the formula: , in, This represents the average standard height. This indicates a ring-shaped neighborhood with a preset radius of 10mm, which limits the search range of normal tissue pixels; This represents the pixel value of the maternal surface mask. A value of 1 indicates that the pixel is located in the placental maternal surface region. as well as express The pixel coordinate variables represent the horizontal and vertical coordinates of the pixel to be examined within the neighborhood, respectively, and are used to traverse all pixels within the annular neighborhood. The pixel value represents the binary defect mask. A value of 0 indicates that the pixel is a normal pixel. This represents the height value corresponding to a normal tissue pixel in the neighborhood. This indicates that the average value is calculated. This formula can be used to obtain the average height of the normal tissue around each defect point, which can be used as a reference for calculating the depth of the depression.

[0046] The difference between the defect height value and the corresponding average standard height value at each defect point is defined as the indentation depth at that point, and the corresponding formula expression is: , in, Indicates the defect point The depth of the depression This represents the average standard height. Indicates the defect point The height value is used to determine the depth of the defect. A positive difference indicates that the defect point is depressed compared to the normal tissue, and the larger the value, the deeper the depression. After traversing the depression depth of all defect points, the maximum value of all depression depths is taken as the maximum depression depth, and the arithmetic mean of all depression depths is taken as the average depression depth. This is used to quantify the overall depression degree of the defect area.

[0047] The geometric centroid coordinates of the defective region are calculated based on the binary defect mask, and the corresponding formula is as follows: , , in, This represents the pixel value of the binary defect mask, where the pixel value is 1 in the defect area and 0 in the non-defect area; The x-coordinate represents the geometric centroid of the defective region. The ordinate represents the geometric centroid of the defective region, which indicates the spatial center of the defective region. The summation symbol is used.

[0048] Based on the geometric centroid coordinates and the preset pallet center coordinates, the spatial position information is determined. First, the distance between the centroid and the pallet center is calculated, and the corresponding formula expression is: , in, This represents the distance from the center of mass to the center of the tray, in centimeters. and These represent the preset horizontal and vertical coordinates of the tray center. The x-coordinate represents the geometric centroid of the defective region. The ordinate represents the geometric centroid of the damaged area; then the angle between the centroid and the center of the tray is calculated using the following formula: ,in, Representing polar coordinate angles, through The function can obtain The angle value between π and π is then converted into a clock face that is easily interpreted clinically, for example, 30° corresponds to the 1 o'clock position, ultimately forming a spatial location description that includes distance and orientation, facilitating precise clinical localization of the defect area. Finally, the defect area is determined based on the maximum indentation depth, average indentation depth, and spatial location information. This requires comprehensive quantitative determination based on preset defect depth and defect area thresholds. The preset depth and area thresholds are calibrated through clinical case statistics of placental remnants, obstetric treatment guidelines, and correlation analysis of late postpartum hemorrhage risk, effectively distinguishing between clinically insignificant physiological micro-indentations and pathological placental defects requiring intervention. After all parameters are calculated, the maximum indentation depth and average indentation depth are compared with the preset depth thresholds, and the defect area corresponding to the binary defect mask is verified against the preset area threshold. When the depth and area parameters of the defect area both meet the preset threshold requirements, and the defect area is accurately located and its contour is locked based on the spatial location information, the area can be identified as a placental defect area with clinical diagnostic significance.

[0049] By calculating the depth of the depression at each point of the defect, the degree of depression in the defect area is accurately quantified, avoiding the limitation of subjective judgment of the depth of depression by manual examination. Based on the spatial location calculation of geometric centroid and polar coordinate transformation, the defect location is transformed into a clinically understandable clockwork orientation and distance description, improving the objectivity and readability of defect localization. The standardized quantitative calculation process provides objective data support for clinical decision-making, and the generated quantitative results can be archived to facilitate medical quality control and dispute evidence, effectively improving the standardization and intelligence level of placental defect detection.

[0050] The above describes the placental defect detection method based on multispectral imaging in the embodiments of the present invention. The following describes the placental defect detection device based on multispectral imaging in the embodiments of the present invention. Please refer to [link to relevant documentation]. Figure 2 One embodiment of the placental defect detection device based on multispectral imaging in this invention includes: Image fusion module 201: used to acquire RGB and infrared images of the placenta, and generate a target fused image based on the RGB and infrared images; Semantic segmentation module 202: used to call a pre-trained semantic segmentation model to perform semantic segmentation processing on the target fused image to obtain a placental maternal surface region image; Defect recognition module 203: used to perform defect recognition processing on the maternal surface area image of the placenta to obtain a binary defect mask; Height map construction module 204: used to construct a height map based on the infrared image; Defect Region Determination Module 205: Used to determine the defect region of the placenta based on the binary defect mask and the height map, and to perform visualization processing on the defect region to obtain a detection report.

[0051] Based on the same ideas as the methods in the above embodiments, the apparatus provided in this application can implement the methods in the above embodiments.

[0052] above Figure 2 The multispectral imaging-based placental defect detection device in this embodiment of the invention is described in detail from the perspective of modular functional entities. The multispectral imaging-based placental defect detection device in this embodiment of the invention is described in detail below from the perspective of hardware processing.

[0053] Figure 3 This is a schematic diagram of a placental defect detection device based on multispectral imaging provided in an embodiment of the present invention. The placental defect detection device 300 based on multispectral imaging can vary considerably due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the placental defect detection device 300 based on multispectral imaging. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the placental defect detection device 300 based on multispectral imaging to implement the steps of the placental defect detection method based on multispectral imaging provided in the above-described method embodiments.

[0054] The placental defect detection device 300 based on multispectral imaging may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The illustrated structure of the placental defect detection device based on multispectral imaging does not constitute a limitation on the placental defect detection device based on multispectral imaging. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0055] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of a placental defect detection method based on multispectral imaging.

[0056] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0057] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0058] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting placental defects based on multispectral imaging, characterized in that, include: Acquire RGB and infrared images of the placenta, and generate a target fusion image based on the RGB and infrared images; The pre-trained semantic segmentation model is invoked to perform semantic segmentation processing on the target fused image to obtain the placental maternal surface region image; The image of the maternal surface region of the placenta is processed for defect identification to obtain a binary defect mask; A height map is constructed based on the infrared image; The defect area of ​​the placenta is determined based on the binary defect mask and the height map, and the defect area is visualized to obtain a detection report.

2. The method for detecting placental defects based on multispectral imaging according to claim 1, characterized in that, The process of generating a target fused image based on the RGB image and the infrared image includes: An image registration algorithm is used to spatially align the RGB image and the infrared image to obtain a dual-channel aligned image. The dual-channel aligned image is fused using a channel stitching and fusion algorithm to obtain an initial fused image. The initial fused image is denoised using a filtering algorithm to obtain a denoised fused image. The denoised fused image is contrast-enhanced using a histogram equalization algorithm to obtain the target fused image.

3. The method for detecting placental defects based on multispectral imaging according to claim 1, characterized in that, The semantic segmentation model includes an encoder module, a decoder module, and a mask generation module, which are sequentially connected. The process of calling the pre-trained semantic segmentation model to perform semantic segmentation on the target fused image to obtain a placental maternal surface region image includes: Based on the encoder module, the target fused image is subjected to multi-scale downsampling feature extraction processing to obtain a compressed semantic feature map; Based on the decoder module, feature fusion processing is performed on the compressed semantic feature map to obtain a high-resolution feature map; The high-resolution feature map is binarized and region extracted based on the mask generation module to obtain the maternal surface region image of the placenta.

4. The method for detecting placental defects based on multispectral imaging according to claim 1, characterized in that, The defect identification processing of the maternal surface region image of the placenta to obtain a binary defect mask includes: Obtain all pixels of the placental maternal surface region image, a predefined mean calculation function, a predefined standard deviation calculation function, a preset grayscale mean threshold, and a preset texture standard deviation threshold; Multiple neighborhood windows of preset sizes are obtained, centered on each of the aforementioned pixels. The local grayscale mean of each neighborhood window is calculated based on a predefined mean calculation function; The local grayscale standard deviation of each neighborhood window is calculated based on a predefined standard deviation calculation function; Each local grayscale mean is compared with the grayscale mean threshold, and each local grayscale standard deviation is compared with the texture standard deviation threshold. Pixels whose local grayscale mean is less than the grayscale mean threshold and whose local grayscale standard deviation is less than the texture standard deviation threshold are identified as defective points, and a binary defect mask is generated based on the defective points.

5. The placental defect detection method based on multispectral imaging according to claim 4, characterized in that, The step of generating the binary defect mask based on the defect points includes: A pixel-level assignment algorithm is used to generate a binary image based on the defect points; The binary image is processed by a morphological opening operation algorithm to obtain a cleaned binary image. The cleaned binary image is processed by a morphological closing operation algorithm to obtain a smoothed binary image. The smooth binary graph is subjected to connected component analysis using a connected component filtering algorithm to obtain the binary defect mask.

6. The method for detecting placental defects based on multispectral imaging according to claim 1, characterized in that, The construction of the height map based on the infrared image includes: Obtain the preset reflectivity, predefined mapping function, and predefined tray plane height reference parameters; A Lambertian reflection model is constructed based on the reflectivity and the infrared image, and the placental surface normal vector is calculated based on the Lambertian reflection model. The placental surface gradient field is calculated using the mapping function based on the placental surface normal vector; An integral algorithm is used to calculate the gradient field on the placental surface to obtain a placental surface height map. The placental surface height map is obtained by performing global calibration processing based on the tray plane height reference parameters.

7. The placental defect detection method based on multispectral imaging according to claim 4, characterized in that, The determination of the placental defect region based on the binary defect mask and the height map includes: Obtain the preset pallet center coordinates, the average standard height, and the defect height value corresponding to each defect point in the binary defect mask; The difference between each defect height value and the average standard height value is defined as the indentation depth of the corresponding defect point, resulting in multiple indentation depths. The maximum indentation depth and the average indentation depth are determined based on the multiple indentation depths. The geometric centroid coordinates are calculated based on the binary defect mask, and the spatial position information is determined based on the distance and angle between the geometric centroid coordinates and the center coordinates of the tray. The defect area is determined based on the maximum indentation depth, the average indentation depth, and the spatial location information.

8. A placental defect detection device based on multispectral imaging, characterized in that, include: Image fusion module: used to acquire RGB and infrared images of the placenta, and generate a target fused image based on the RGB and infrared images; Semantic segmentation module: used to call a pre-trained semantic segmentation model to perform semantic segmentation processing on the target fused image to obtain the placental maternal surface region image; Defect recognition module: used to perform defect recognition processing on the maternal surface area image of the placenta to obtain a binary defect mask; Height map construction module: used to construct a height map based on the infrared image; Defect Region Determination Module: Used to determine the defect region of the placenta based on the binary defect mask and the height map, and to visualize the defect region to obtain a detection report.

9. A placental defect detection device based on multispectral imaging, characterized in that, The placental defect detection device based on multispectral imaging includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the multispectral imaging-based placental defect detection device to perform the steps of the multispectral imaging-based placental defect detection method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the placental defect detection method based on multispectral imaging as described in any one of claims 1-7.

Images