A multi-modal data-driven based method for rapid identification of microplastic components

Abstract

The application relates to the technical field of computer vision and environment monitoring, and discloses a microplastic component rapid identification method based on multi-modal data driving, which comprises the following steps: firstly, acquiring morphological sequence images and chemical component maps of microplastics, and constructing a standardized three-dimensional voxel dataset through metadata alignment; then, constructing a non-coupling training set and performing three-dimensional data enhancement based on spatial transformation; training a component identification network by using the enhanced data, and establishing a mapping relationship between the three-dimensional morphological characteristics of the microplastics and the chemical components; finally, collecting morphological images of samples to be measured and inputting the images into the network, and outputting three-dimensional segmentation masks with chemical component codes. The application realizes rapid and accurate identification of the chemical components and three-dimensional distribution of microplastics only according to low-cost morphological images by mining the deep connection between the morphology and the components, solves the problems of high cost and slow speed of traditional chemical imaging, and is suitable for efficient monitoring of environmental microplastic pollution.

Description

Technical Field

[0001] This invention relates to the field of computer vision and environmental monitoring technology, specifically to a rapid identification method for microplastic components based on multimodal data-driven approaches. Background Technology

[0002] Currently, microplastics, as an emerging pollutant, are widely present in water bodies, soil, and biological media. These particles are extremely small in size and have diverse chemical compositions. Environmental monitoring requires accurate identification of their polymer types and physical forms. Standard testing procedures typically include environmental sample collection, followed by pretreatment steps such as organic matter digestion and density flotation. The target particles are ultimately enriched onto a filter membrane for subsequent analysis. This process forms the data foundation for pollution source tracing and ecological risk assessment.

[0003] For the enriched particulate samples described above, existing techniques primarily employ physicochemical analysis. Microscopic Fourier transform infrared spectroscopy and microscopic Raman spectroscopy are widely used, both identifying polymer types by capturing the molecular vibrational fingerprint spectrum of the material. Laser direct infrared imaging utilizes quantum cascade lasers to perform chemical imaging scans of the samples. Furthermore, pyrolysis gas chromatography-mass spectrometry analyzes the chemical composition of the polymer matrix by determining the pyrolysis products after high-temperature combustion of the sample.

[0004] However, the aforementioned technologies have limitations in practical applications. Spectroscopic imaging devices often employ point-by-point or line-by-line scanning modes, resulting in long acquisition cycles for full-filter data. This time cost limits the efficiency of high-throughput sample monitoring. High-precision spectroscopic analysis instruments typically rely on sophisticated optical systems, placing specific requirements on the operator's spectral interpretation skills and the equipment maintenance environment. While ordinary optical microscopes acquire images quickly, they can only record superficial morphological information such as color and shape. Due to the lack of chemical dimension characteristic data, it is difficult to distinguish between transparent mineral particles and microplastics, or accurately identify chemically similar components such as polyethylene and polypropylene, relying solely on optical images.

[0005] Therefore, this invention provides a rapid identification method for microplastic components based on multimodal data to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a rapid identification method for microplastic components based on multimodal data-driven methods. This method solves the technical problems of long detection time and high cost of spectral analysis equipment in existing microplastic detection technologies, and low identification accuracy caused by the inability of ordinary optical microscopes to identify chemical components.

[0007] This invention provides a rapid identification method for microplastic components based on multimodal data, employing the following technical solution:

[0008] A rapid identification method for microplastic components based on multimodal data-driven methods includes the following steps:

[0009] Morphological sequence images and chemical composition maps of microplastics are obtained. The chemical composition maps are parsed and mapped into three-dimensional voxel-level composition labels. A standardized three-dimensional voxel dataset containing morphological features and composition labels is constructed through physical space consistent metadata alignment.

[0010] Based on the aforementioned 3D voxel dataset, a training set decoupled from the morphological image list and the component label list is constructed, and 3D data augmentation based on spatial transformation is performed.

[0011] A component recognition network was trained using a training set augmented with spatial transformation-based 3D data to establish a mapping relationship between the 3D morphological features of microplastics and the distribution of chemical components.

[0012] Morphological sequence images of the sample to be tested are acquired and input into the trained component recognition network. The mapping relationship is used to output a three-dimensional segmentation mask with chemical component encoding, thereby realizing the rapid identification of microplastic components.

[0013] By employing the above technical solution, and utilizing deep learning networks to mine and establish the potential nonlinear mapping relationship between the external morphology and internal chemical composition of microplastics, in practical applications, only low-cost and fast morphological sequence images need to be input, eliminating the need for time-consuming chemical spectral scanning, to directly infer the three-dimensional chemical composition distribution of the sample. This asymmetric mode of multimodal input in the training phase and single-modal input in the inference phase solves the bottleneck problems of slow detection speed and low throughput of traditional infrared or Raman spectroscopy imaging, improving the efficiency and real-time performance of microplastic detection while maintaining the accuracy of chemical identification.

[0014] Preferably, the process of constructing a standardized three-dimensional voxel dataset through physical space-consistent metadata alignment includes: grouping the original two-dimensional slice sequences of the morphological sequence image using unique identifiers for sequence instances to identify slice sets belonging to the same microplastic sample; stacking the slice sets into a three-dimensional voxel array along a direction perpendicular to the image plane based on the position information recorded in the metadata; converting the three-dimensional voxel array into NIfTI format and extracting the origin coordinates, orientation matrix, and voxel spacing parameters from the header information of the morphological sequence image; parsing the chemical composition map and generating a corresponding three-dimensional label file, and forcibly writing the origin coordinates, orientation matrix, and voxel spacing parameters into the corresponding three-dimensional label file to achieve alignment of the morphological sequence image and the chemical composition map in the physical coordinate system.

[0015] By adopting the above technical solution, physical spatial parameters are encapsulated using a common data format standard for medical imaging, thus forcibly unifying the coordinate system definitions of morphological data and chemical spectral data. This eliminates spatial positional deviations caused by differences in sampling resolution, scanning center, or scanning direction between the two imaging devices, ensuring that the morphological features of each voxel point accurately correspond to the chemical label, and providing high-quality basic data support for training a high-precision component recognition network.

[0016] Preferably, the process of constructing a training set that is decoupled between the morphological image list and the component label list includes: configuring a keyword list containing preset feature strings, wherein the preset feature strings are used to identify file attributes containing true chemical component information; traversing data storage paths, writing file paths whose filenames contain the preset feature strings into the component label list, and writing file paths whose filenames do not contain the preset feature strings into the morphological image list; reading sample metadata attributes, and stratifying them based on microplastic polymer type or morphological category to ensure that the proportion of each category in the training set is consistent with that in the validation set; initializing a random number generator with fixed random seed parameters to generate a set of determined index permutation vectors; using the index permutation vectors to perform synchronous rearrangement of the morphological image list and the component label list, and truncating the training set according to a preset ratio.

[0017] By adopting the above technical solution, the logical decoupling of data index and physical files is achieved, avoiding resource overflow caused by directly loading large-scale 3D data into memory. Simultaneously, the synchronous rearrangement mechanism of hierarchical sampling and fixed random seeds ensures statistical consistency in sample class distribution between the training and validation sets, effectively preventing overfitting or evaluation bias in the network model due to uneven data partitioning.

[0018] Preferably, the process of performing 3D data augmentation based on spatial transformation includes: constructing a four-dimensional affine transformation matrix that defines the mapping relationship from the original voxel coordinate system to the target augmentation coordinate system; randomly generating scaling coefficients along the three axial directions, and using the four-dimensional affine transformation matrix to calculate the new position of the voxel in space; for the morphological sequence image data in the training set, using a trilinear interpolation algorithm to calculate the gray value at the new position; for the label data corresponding to the chemical composition maps in the training set, using a nearest neighbor interpolation algorithm to select the original voxel label value with the closest geometric distance as the label value at the new position. More preferably, the process further includes: calculating all permutation orders along the three spatial axes and the positive and negative reading states of each axis based on the geometric characteristics of the three-dimensional Cartesian coordinate system; combining the six axial sequences and eight directional states to generate forty-eight unique spatial dimension permutation key values; randomly selecting one of the spatial dimension permutation key values, and performing a dimension transpose operation and a reverse operation on a preset dimension on the samples in the training set; and simultaneously applying the dimension transpose operation and the reverse operation to the morphological input image volume and the chemical composition label volume.

[0019] By employing the aforementioned technical solution, and considering the random orientation and diverse morphologies of microplastic particles in the natural environment, affine transformation and omnidirectional spatial dimensional rearrangement were used to simulate various posture changes of samples in three-dimensional space. For the different attributes of image data and discrete label data, smooth trilinear interpolation and value-preserving nearest-neighbor interpolation were employed respectively, expanding the data scale while maintaining the accuracy of label categories. This enables the network model to learn robust features with rotation invariance and scale invariance, solving the training difficulty problem caused by the scarcity of three-dimensional labeled data.

[0020] Preferably, the process of training the component recognition network using the enhanced training set includes: constructing a dynamic U-Net network architecture containing an encoder, a decoder, and skip connections as the component recognition network; extracting the geometric and texture features of microplastics using multi-layer convolutional modules in the encoder, and gradually reducing the spatial resolution of the feature maps by setting the convolution stride; upsampling the deep feature maps using the decoder through a three-dimensional transposed convolution operation, and concatenating the corresponding feature maps output by the encoder through the skip connections; mapping the high-dimensional features to the corresponding chemical component category probability maps using the convolutional operation of the output layer, and updating the parameters using the DiceLoss loss function combined with the Adam optimizer. The process of updating parameters using the DiceLoss loss function combined with the Adam optimizer includes: applying the Sigmoid activation function to the feature map of the output layer of the component recognition network to obtain the probability value of voxels belonging to a preset chemical component category; calculating the intersection between the predicted volume and the true label volume, and calculating the Dice loss value based on the ratio of the intersection to the sum of the predicted volume and the true label volume; dynamically adjusting the learning rate using the Adam optimizer based on the first moment estimate and the second moment estimate of the parameters; monitoring the component recognition accuracy on the validation set, and triggering a learning rate decay mechanism when the component recognition accuracy does not improve within a preset period.

[0021] By employing the above technical solution, the skip connection structure of the U-Net architecture effectively integrates deep semantic features with shallow spatial details, improving the segmentation capability for the edges of tiny plastic particles. In particular, the Dice loss function based on set similarity directly optimizes the spatial overlap between the predicted and real regions, effectively alleviating the class imbalance problem in microplastic detection scenarios where the number of background voxels far exceeds the number of target voxels, and preventing the network from favoring the prediction of the background while ignoring fine particles.

[0022] Preferably, establishing the mapping relationship between the three-dimensional morphological features of microplastics and the distribution of chemical components includes setting up a hierarchical labeling system. The specific construction steps of the labeling system are as follows: defining a first-level label, which is used to distinguish the microplastic foreground from the environmental background; defining a second-level label, which classifies the microplastic foreground into polyolefins and non-polyolefins based on chemical structural similarity; defining a third-level label, which uses unique thermal coding technology to identify specific polymer types, including polyethylene, polypropylene, and polystyrene; and driving the component recognition network to output the predicted probabilities of the first-level label, the second-level label, and the third-level label at the output end.

[0023] By employing the above technical solution, and through hierarchical supervision signals ranging from coarse to fine, the network model is guided to prioritize learning the obvious differences between microplastics and the background, then learn the common characteristics of different polymer groups, and finally focus on the subtle differences in specific chemical components. This lesson-based learning strategy reduces the difficulty of model convergence and improves the classification accuracy for plastics with similar chemical structures.

[0024] Preferably, the specific steps for outputting the 3D segmentation mask with chemical composition encoding are as follows: preprocessing the morphological sequence image of the sample to be tested into a 3D voxel matrix consistent with the format of the standardized 3D voxel dataset; inputting the preprocessed 3D voxel matrix into the composition recognition network to obtain a multi-channel probability feature map; performing a maximum value indexing operation on the channel dimension of the probability feature map, and using the channel index with the maximum probability value as the predicted category label of the voxel; generating the 3D segmentation mask with chemical composition encoding, where the background region is labeled as zero and the microplastic region is labeled as the integer encoding of the corresponding polymer type. The method also includes generating an environmental monitoring report based on the 3D segmentation mask with chemical composition encoding. The specific process is as follows: extracting the physical spacing parameters of the 3D voxel matrix to be measured in three orthogonal directions, and calculating the physical volume of a single voxel; counting the total number of voxels under each polymer category, and calculating the absolute volume and proportion of each microplastic component; performing connected component analysis on the 3D segmentation mask with chemical composition encoding using the 26-neighborhood connectivity rule to identify independent microplastic particles; calculating the maximum and minimum Feretta diameters of the independent microplastic particles, and classifying the independent microplastic particles into fibrous, fragmented, or spherical shapes based on the aspect ratio; and summarizing the absolute volume and proportion of each microplastic component with the classification results of the independent microplastic particles to generate the environmental monitoring report.

[0025] By adopting the above technical solution, the entire process from raw image input to final environmental monitoring report output has been automated. It can not only identify the chemical composition of microplastics, but also accurately calculate the absolute volume, particle size distribution, and morphological type of particles based on the physical properties of the three-dimensional voxel matrix being measured, providing multi-dimensional quantitative data support for environmental microplastic pollution assessment.

[0026] This invention provides a rapid identification method for microplastic components based on multimodal data-driven approaches. It offers the following advantages:

[0027] 1. This invention achieves a paradigm shift in detection methods by establishing a nonlinear mapping relationship between the three-dimensional morphological features of microplastics and their chemical composition. In the application stage, this method only requires acquiring low-cost morphological images to infer chemical composition using a trained component recognition network, eliminating the need for expensive Fourier transform infrared spectrometers or Raman spectrometers for point-by-point scanning. This decoupled architecture of multimodal fusion at the training end and single-modal inference at the application end reduces the hardware cost and equipment threshold for microplastic monitoring, enabling ordinary microscopic imaging equipment to identify chemical components.

[0028] 2. This invention leverages the parallel computing capabilities of deep learning networks to improve detection speed and throughput. Compared to the time-consuming point-by-point or line-by-line scanning imaging of traditional spectroscopic techniques, this invention, by inputting a standardized three-dimensional voxel dataset, can rapidly output a three-dimensional segmentation mask with chemical composition encoding within seconds to minutes using a component recognition network. This end-to-end rapid recognition mechanism eliminates the physical bottleneck of spectral acquisition, improves the efficiency of large-scale environmental sample surveys, and achieves a leap from detailed laboratory analysis to high-throughput monitoring.

[0029] 3. This invention effectively overcomes the subjectivity and limitations of traditional manual microscopic examination by employing 3D data augmentation based on spatial transformation and a dynamic U-Net network architecture. The model, trained in an enhanced high-dimensional feature space, can uncover potential morphological features that are intangible to the human eye, such as fine textures and differences in edge refractive index, thereby accurately establishing their correspondence with polymer types. This not only solves the technical challenge of distinguishing chemical components based solely on morphology but also reduces the false positive rate caused by subjective misjudgment, achieving objectivity and standardization in microplastic component detection. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall process of the rapid identification method for microplastic components based on multimodal data driven by the present invention;

[0031] Figure 2 This is a schematic diagram of the microplastic component identification network architecture and training process based on DynUNet of the present invention;

[0032] Figure 3 This is a schematic diagram of the standardized three-dimensional voxel grayscale feature distribution simulation of the present invention;

[0033] Figure 4 This is a schematic diagram comparing the training convergence performance of the component recognition network of the present invention. Detailed Implementation

[0034] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] This invention provides a method for rapid identification of microplastic components based on multimodal data. This method utilizes deep learning networks to mine the mapping relationship between the morphological features and chemical composition of microplastics in three-dimensional space. High-precision spectral data or microscopic CT data are used as supervisory signals, and the trained model can infer the chemical composition and three-dimensional distribution of microplastics based solely on low-cost optical or morphological images. The method in this embodiment is based on computer vision and deep learning technologies, establishing a complete technical process from multimodal data acquisition, three-dimensional normalization reconstruction, decoupled data partitioning, spatial augmentation to fully automated component inference.

[0036] This invention can be executed on a computer device containing a high-performance computing unit. The computer device is equipped with a central processing unit (CPU) and a graphics processing unit (GPU), where the GPU is used to accelerate parallel computation of three-dimensional convolutional neural networks. In a specific implementation environment, a server equipped with an NVIDIA series graphics card is used as the computing platform, with memory configuration sufficient to process high-resolution three-dimensional voxel data. At the data acquisition end, the imaging equipment involved includes an optical microscope or a micro-computed tomography scanner with Z-axis scanning capability for acquiring morphological sequence images of microplastics; and a laser direct infrared imager or a micro-infrared spectrometer for acquiring chemical composition maps of microplastics. The software environment is based on the Python programming language, utilizing the PyTorch deep learning framework and the MONAI medical image processing framework to construct the algorithmic logic for model training and inference.

[0037] See attached document Figure 1 The method first executes step S1, which acquires multimodal data and reconstructs three-dimensional volume. In this step, the system acquires morphological sequence images and chemical composition maps of microplastic samples. Due to differences in the output format and spatial resolution of different imaging devices, the system uses unique identifiers for sequence instances to group the original DICOM or TIFF sequences, uses the ITK library to reconstruct the two-dimensional slice sequences into three-dimensional volume data, and uniformly converts the morphological data and chemical labels into the NIfTI format, which includes spatial coordinate information and voxel spacing, thereby achieving strict alignment of the morphological data and chemical labels in physical space and constructing a standardized three-dimensional voxel dataset.

[0038] After data standardization, the method executes step S2 to construct and partition a decoupled intelligent dataset. The system employs a keyword-based feature separation indexing mechanism, defining a list of tag keywords containing specific identifiers. It automatically traverses the storage directory to distinguish between morphological image files and component label files, without relying on traditional one-to-one filename matching. Subsequently, the system stratifies data according to the category or component category of microplastics and sets a fixed random seed. This ensures that the morphological image list and the component label list, while logically independent, can be divided into training and validation sets according to the same random distribution ratio, thereby preventing data leakage and ensuring the diversity of training data.

[0039] Next, the method executes step S3, performing 3D data augmentation based on spatial transformation. Considering the diverse morphologies and random orientations of microplastic particles in the medium, the system simultaneously applies affine matrix transformation to the input microplastic volume data and its corresponding chemical labels. Resampling techniques are used to simulate imaging differences under different microscopic magnifications and focal lengths. Simultaneously, the system generates multiple spatial dimension arrangement key values, performing real-time axial rearrangement and rotation operations on the training samples to simulate arbitrary deposition angles of microplastics in water or filter membranes. This forces the deep learning network to learn rotationally invariant chemical texture features, and the input data is randomly cropped and padded to fit the network input size.

[0040] Subsequently, the method executes step S4, constructing and training a component recognition network based on DynUNet. (See Appendix) Figure 2 The system constructs a three-dimensional convolutional neural network containing a dynamic encoder, decoder, and skip connections. It progressively extracts the macroscopic geometric features and microscopic texture features of microplastics through multi-layer stride convolutions. During training, the system sets up a hierarchical labeling system, first identifying the background and foreground, then further subdividing into specific polymer types. It uses the DiceLoss loss function combined with the Adam optimizer for iterative updates of model parameters, and employs a learning rate decay strategy to find the global optimum, thereby establishing a nonlinear mapping model from morphological features to chemical composition.

[0041] Finally, the method executes step S5 for practical inference and detection applications. The user acquires morphological images of the sample using a conventional optical microscope or micro-CT. The system automatically performs preprocessing and NIfTI format conversion, loads the trained model for forward inference, and outputs a 3D segmentation mask with chemical composition encoding. Based on the segmentation results, the system further calculates statistical indicators such as the volume, number, and aspect ratio of each component particle, generating an environmental monitoring report containing the proportion and distribution of microplastic components.

[0042] In step S1, acquiring multimodal data sources is the foundation for constructing the training dataset, mainly including the acquisition of morphological input data and chemical label data.

[0043] For the acquisition of morphological input data, this embodiment of the invention utilizes a high-resolution three-dimensional imaging device to obtain internal texture and surface morphology information of microplastic samples. Specifically, a filter membrane or substrate containing a microplastic sample is placed on the stage of a computed tomography (CT) scanner or an optical microscope with Z-axis scanning capability. The imaging device scans the sample layer by layer along the Z-axis direction perpendicular to the filter membrane plane at a preset interlayer spacing. During the scanning process, the device records the light intensity or density information of the sample at different depth levels, generating a series of continuous two-dimensional slice images. These two-dimensional slice images are arranged in spatial depth order, constituting a morphological sequence image containing the three-dimensional spatial structure of the microplastics. The data format of this morphological sequence image is stored as a sequence file in medical digital imaging communication format or labeled image file format, where each frame corresponds to a specific cross-section of the sample, recording the edge geometric features, surface roughness features, and internal cracks or voids caused by aging of the microplastic particles.

[0044] For the acquisition of chemically labeled data, this embodiment of the invention utilizes a spectral analysis device to obtain the exact category of chemical components in the same sample area, which serves as the ground truth label for supervised learning. While maintaining the sample position relative to the morphological acquisition stage or repositioning it via coordinate marking, a laser direct infrared imager or a micro-Fourier transform infrared spectrometer is used to perform a full-plane spectral scan of the sample. The spectral analysis device acquires the absorption or reflection spectra of microplastic particles in specific infrared bands. By comparing these spectra with a standard polymer spectral library, the chemical component category corresponding to each pixel is determined, generating a two-dimensional chemical component distribution map. Different pixel values ​​in this chemical component distribution map represent different polymer types, such as polyethylene, polypropylene, or polystyrene.

[0045] To achieve dimensional matching between chemical tags and three-dimensional morphological data, this embodiment of the invention maps a two-dimensional chemical composition distribution map to voxel tags in three-dimensional space. Since microplastic particles are typically composed of a single material, their internal chemical composition is homogeneous. The system projects and extends the two-dimensional chemical composition distribution map along the Z-axis. Specifically, the system identifies the three-dimensional voxel region occupied by the microplastic particles in the morphological sequence image and assigns the corresponding two-dimensional spectral classification result to all voxels within that three-dimensional region. In this way, three-dimensional tag volume data that completely corresponds to the morphological sequence image in spatial dimension is generated. The value of each voxel in this three-dimensional tag volume data is no longer grayscale or density, but rather the chemical composition category code of the substance at that location.

[0046] In step S1, volume reconstruction based on sequence metadata is a crucial step in converting discrete two-dimensional image data into computer-processable three-dimensional voxel data. This embodiment of the invention employs an automated script program to logically group and spatially reconstruct massive amounts of raw scan files.

[0047] First, the system performs sequence grouping based on the unique identifier of the sequence instance. Since microcomputed tomography scanners or optical microscopes typically store all sample slice images together in the same output directory during scanning, or although stored in separate folders, they lack a unified index, the system needs to automatically identify which slices belong to the same scan of the same microplastic sample. The system traverses the specified data storage path, reading the header metadata information of all image files in each directory. The system extracts the unique identifier of the sequence instance from the metadata; this field is a globally unique code used in medical digital imaging communication standards to distinguish different scan sequences. The system maps all 2D slice files with the same unique identifier of the sequence instance to a separate list of slice sets, thereby achieving logical isolation of different microplastic sample data or different scan batches of data, ensuring that subsequent 3D reconstruction does not result in data confusion between samples.

[0048] Secondly, the system performs a physical reconstruction process from 2D slices to 3D volume data. For each grouped slice set list, the system calls the sequence reading function in the image processing library for processing. The system first reads the image location information or slice instance number recorded in the metadata of each 2D slice in the set. Based on the value of this location information in the Z-axis direction, the slice file paths are strictly sorted in ascending or descending order to restore the correct hierarchical order of the slices in physical space. Subsequently, the system loads the pixel matrix data of each 2D slice sequentially according to the sorted list, and stacks these 2D matrices along the third dimension perpendicular to the image plane to construct a continuous 3D voxel array. During the stacking process, the system simultaneously reads the pixel spacing, layer thickness, and image origin coordinates recorded in the metadata, and assigns these spatial parameters to the generated 3D voxel array. This not only restores the geometry of the microplastics but also preserves their true physical size proportions, providing an accurate measurement benchmark for subsequent calculations of the volume and surface area of ​​the microplastics.

[0049] In step S1, after constructing the three-dimensional voxel array, the system performs NIfTI format standardization and spatial alignment operations. This embodiment of the invention selects the NIfTI format with compressed encoding as the standard container for data storage. The necessity of this format conversion lies in the fact that the original two-dimensional slice sequence or simple binary volume data cannot carry the location and scale information of the sample in real physical space. By converting to NIfTI format, the system simultaneously encapsulates the three-dimensional voxel grayscale matrix and metadata describing spatial geometric properties in a single file structure. The core of this metadata includes the affine transformation matrix and the voxel spacing parameter. The affine transformation matrix defines the linear mapping relationship from the discrete voxel index coordinate system to the continuous physical world coordinate system, clarifying the origin position and direction cosine of the image. The voxel spacing parameter accurately records the physical sampling resolution of the microplastic sample in the three Cartesian coordinate axes, ensuring the dimensional accuracy when subsequently calculating the geometric features such as the volume and surface area of ​​the microplastic.

[0050] To ensure strict alignment between morphological data and chemical labels in the physical coordinate system, the system implements a metadata inheritance strategy based on physical spatial consistency. When generating 3D voxel labels for chemical components, the system directly reads the header information of the corresponding morphological structure image, extracting the origin coordinates, orientation matrix, and voxel spacing values. The system then forcibly writes these extracted spatial parameters into the header metadata of the chemical label file, ensuring that the chemical label file and the morphological structure image file share the exact same physical coordinate definition. This forced alignment mechanism guarantees that each voxel in the morphological image and the voxel at the same index position in the chemical label image represent the same microscopic region of the microplastic in physical space. This strict correspondence in physical space eliminates the risk of feature misalignment caused by coordinate system mismatch, achieving precise fusion of low-cost morphological features and high-precision chemical composition features in three-dimensional space.

[0051] See attached document Figure 1 This invention provides a decoupled intelligent dataset construction and partitioning method, which specifically includes a keyword-based feature separation indexing process in step S2. This process aims to automatically identify and separate hybrid morphological structure image files and chemical composition label files through predefined semantic rules, thereby eliminating rigid dependence on specific file naming formats.

[0052] The system first configures a keyword list containing predefined feature strings, which is stored in the system's configuration memory. The feature strings are specifically designed to identify file attributes containing true information about chemical composition. In one specific embodiment of the invention, the keyword list includes the strings `seg`, `label`, `mask`, and `composition`. These strings represent semantic identifiers such as segmentation results, annotation information, mask data, and component distribution data, respectively, used to indicate file content attributes at the filename level. Simultaneously, the system initializes two independent index containers in memory, defined as a morphological image list and a component label list, for subsequent storage of file access paths after classification and discrimination.

[0053] Next, the system performs a directory traversal operation, recursively scanning the data storage path after the standardization process in step S1. For each NIfTI format file existing in the path, the system reads its filename and converts it into a string. The system then performs an inclusion match detection between this filename string and all feature strings in the keyword list. During the detection process, the processor executes string comparison logic. If the filename string contains any feature string from the keyword list, the processor generates a label determination signal and appends the complete access path of the file to the component label list. If, after a complete comparison, the filename string is confirmed not to contain any feature string from the keyword list, the processor generates an image determination signal and appends the complete access path of the file to the morphological image list.

[0054] Through the keyword-based feature separation indexing mechanism described above, the system achieves logical decoupling between morphological data and chemical label data. This decoupling mechanism allows the system to accommodate raw data from various sources and with different naming conventions. As long as the filename of the component label file contains any key character indicating its attribute, it can be correctly identified and classified by the system. The system does not need to batch rename files or enforce strict character-level correspondence in filenames. Instead, it automatically separates the mixed-storage file stream into a clean morphological input stream and a chemical label stream through semantic feature matching. This process ensures accurate logical separation between morphological image data and chemical component label data, providing a structured and non-interfering data foundation for subsequent dataset partitioning and model training.

[0055] See attached document Figure 1This invention provides a uniformly distributed random partitioning strategy, specifically including a process of hierarchical sampling and deterministic random allocation of morphological image data and component label data in step S2. After completing the feature separation index, the system first executes hierarchical logic based on microplastic categories to ensure consistency in sample distribution between the training and validation sets. The system reads the metadata attributes of each sample and divides the dataset into several mutually exclusive subsets, each subset corresponding to a specific microplastic polymer type or morphological category. This hierarchical mechanism forces that each chemical component category occupy the same proportion in the final partitioned training and validation sets as it does in the overall samples, thereby preventing the model from losing its ability to identify rare microplastic components due to sample category imbalance.

[0056] During the specific partitioning operation, the system introduces a fixed random seed parameter to control the initialization state of the pseudo-random number generator. In this embodiment, the random seed parameter is set to a constant integer value, such as forty-two. This constant value is not intended to introduce randomness, but rather to eliminate uncertainty in the random process. When the system uses this fixed random seed to initialize the random number generation algorithm, the generated pseudo-random number sequence remains unique and unchanged regardless of how many times the program runs. The system uses this determined random sequence to generate a set of index permutation vectors, and applies this same set of permutation vectors to the rearrangement operations of the morphological image list and the component label list, respectively.

[0057] Through the above mechanism, the system logically ensures that the morphological image list and the component label list undergo the exact same disordered transformation process. Since the morphological image list and the component label list are initially arranged according to the lexicographical order of filenames or other deterministic rules, and both follow mathematically identical permutation paths during the reordering process, each element in the reordered morphological image list maintains a strict one-to-one correspondence with its corresponding element in the component label list. This synchronous reordering strategy based on a fixed seed allows the system to maintain the pairing relationships of cross-modal data without establishing a complex key-value pair mapping table, effectively reducing memory overhead and simplifying the data management process.

[0058] After completing the synchronous rearrangement, the system performs a truncation operation according to a preset partitioning ratio, dividing the rearranged list into training and validation subsets. The system then performs physical isolation, moving the morphological structure image files and chemical composition label files belonging to the training subset to a dedicated training directory, and moving the files belonging to the validation subset to a dedicated validation directory. This physical isolation ensures that the validation set data is completely invisible to the neural network during the model training phase, fundamentally preventing data leakage and guaranteeing the authenticity and objectivity of the model performance evaluation results.

[0059] See attached document Figure 1 This invention provides a three-dimensional spatial data enhancement strategy, which specifically includes a three-dimensional affine transformation process in step S3. This process aims to simulate the imaging differences of different optical microscopes or CT devices under different magnification and focal length settings through mathematical spatial mapping transformation, thereby expanding the diversity of samples at the data level and enabling the model to adapt to the changes in the scale of microplastic particles in actual detection.

[0060] The system first loads the 3D volume data of the microplastic to be processed and its corresponding chemical composition label data. To achieve isotropic or anisotropic scaling, the system constructs a four-dimensional affine transformation matrix. This matrix defines the mapping from the original voxel coordinate system to the target augmented coordinate system. The processor applies the following core affine transformation formula to calculate the new position of each voxel in space:

[0061] ;

[0062] In this formula, , , This represents the original spatial coordinates of any voxel in the original three-dimensional volume data. , , This represents the target space coordinates after the transformation. In the matrix... This represents the scaling factor along the X-axis. This represents the scaling factor along the Y-axis. This represents the scaling factor along the Z-axis. The system generates these factors within a preset scaling range using a random sampling algorithm. , , When the value is greater than 1, it simulates a microscopic magnification effect; when the value is less than 1, it simulates a microscopic reduction effect.

[0063] After determining the new spatial coordinate grid, the system must calculate the voxel values ​​at the new grid points. Since the positions after coordinate transformation usually do not fall on integer grid points, the system employs two different resampling interpolation strategies for morphological image data and chemical composition label data.

[0064] For morphological image data, the system employs a trilinear interpolation algorithm. Specifically, for transformed non-integer coordinate points ( , , ), its grayscale value It is obtained by weighted summation of the gray values ​​of its eight surrounding integer coordinate points, and the calculation formula is as follows:

[0065] ;

[0066] In the above formula, , They are Integer coordinates for rounding down and rounding up (i.e.) , Similarly , These are the adjacent integer coordinates along the corresponding axis. This represents the grayscale values ​​of the eight adjacent voxels in the original image. The formula uses the normalized distance from the target point to the eight vertices as weights to achieve a smooth transition of grayscale values ​​in three-dimensional space, ensuring the continuity and smoothness of the internal texture and edge diffraction features of the microplastic particles, and avoiding artificial artifacts caused by scaling.

[0067] This interpolation method can maintain the continuity and smoothness of the internal texture and edge diffraction features of microplastic particles, avoiding artificial artifacts caused by scaling.

[0068] For chemical composition label data, the system employs the nearest neighbor interpolation algorithm. This algorithm directly selects the label value of the original voxel that is geometrically closest to the target coordinate point as the label value of the new grid point, without performing any weighted average calculation. This strategy is adopted because chemical composition labels belong to discrete category data; for example, value 1 represents polyethylene, and value 2 represents polypropylene. Linear interpolation might produce non-integer intermediate values, leading to incorrect category assignment. Nearest neighbor interpolation strictly preserves the original integer category encoding, ensuring the uniqueness and accuracy of microplastic chemical composition attributes during spatial transformation, thereby achieving strict synchronization and alignment between morphological data and label data under geometric deformation.

[0069] See attached document Figure 1 This invention provides a spatial dimension arrangement and attitude simulation strategy, which specifically includes generating a full-coverage spatial transformation key value and performing axial rearrangement operations in step S3. The core objective of this strategy is to construct all possible orthogonal observation attitudes of microplastic particles in three-dimensional space through mathematical means, thereby eliminating the interference of imaging angle on the accuracy of chemical composition identification.

[0070] The system first constructs a set of spatial transformation key values ​​based on the geometric characteristics of a three-dimensional Cartesian coordinate system. For any given three-dimensional voxel data of microplastics, there are three orthogonal spatial axes, defined as the first, second, and third axes. The system calculates all possible permutations of these three spatial axes. According to the principles of permutation mathematics, the three elements generate a total of six different axis sequences. Based on this, the system further calculates the positive and negative read states for each axis. Since each of the three axes has two independent states (positive and negative), there are a total of eight possible combinations of axial directions. The system performs a Cartesian product operation on the six axis sequences and the eight direction states, thereby generating a total of forty-eight unique spatial dimension permutation key values. Each key value represents a specific three-dimensional spatial view transformation logic, covering all possible combinations of 90-degree rotations and mirror flips of the object in three-dimensional space.

[0071] During the real-time augmentation phase of the training data, the system randomly selects a key value from the aforementioned forty-eight spatial dimension arrangements and applies it to the current microplastic sample. Based on the selected key value, the system first performs a dimension transpose operation on the three-dimensional voxel matrix, changing the storage dimension order of the data to simulate the rotation effect of microplastic particles around the diagonal. Next, according to the direction indication in the key value, the system performs a reverse operation on the voxel indices in a specific dimension, achieving a mirror flip along a specific axis. This series of transformation operations is applied simultaneously and consistently to the morphological input image volume and the chemical composition label volume, ensuring that the transformed morphological texture and internal chemical composition distribution maintain precise voxel-level alignment in the new spatial coordinate system.

[0072] This spatial rearrangement mechanism, based on 48 possible permutations, physically simulates the random deposition behavior of microplastic particles during water settling or membrane filtration. In real-world microscopic imaging scenarios, the angle of microplastic particles relative to the imaging sensor is completely random and uncontrollable. By forcing the network to contact the same microplastic sample in 48 different spatial poses during training, the deep learning model is endowed with rotation and flip invariance through a data-driven approach. The model is forced to ignore the absolute spatial orientation of the particles and instead focus on extracting essential features that do not change with spatial pose, such as the topological structure of weathering cracks, the statistical distribution of surface roughness, and the order features of edge grating diffraction, thereby improving the model's robustness in recognizing microplastic samples at arbitrary angles in the real world.

[0073] See attached document Figure 2This invention provides a dynamic U-Net network architecture specifically configured to process preprocessed and normalized 3D microplastic voxel data, aiming to establish a voxel-level mapping relationship between the 3D morphological features of microplastics and their internal chemical composition. The main network structure includes a shrinking path (encoder) for feature extraction and an expanding path (decoder) for restoring spatial resolution, with the two performing feature fusion through skip connections.

[0074] The network input layer receives single-channel density-grayscale volumetric data or multi-channel multispectral volumetric data. The encoder consists of five consecutive cascaded convolutional modules, each containing two 3D convolutional layers followed by a normalization layer and an activation function layer. To balance feature extraction of both microscopic textures and macroscopic geometry on the microplastic surface, the encoder employs a specific convolutional stride configuration strategy. The first convolutional module has a stride of 1 in all three dimensions (1, 1, 1). This configuration ensures that the spatial resolution of the feature map is not reduced in the first layer of the network, thus fully preserving the subtle weathering cracks, edge roughness, and high-frequency texture information of the microplastic surface. Starting with the second convolutional module, the subsequent three modules have a stride of 2 in all three dimensions (2, 2, 2). This progressive downsampling design halves the spatial size of the feature map in each of the length, width, and height dimensions after each pass, thereby gradually expanding the receptive field.

[0075] As spatial resolution decreases, the number of filters in each layer of the network increases exponentially. The first layer has 32 filters, which increases to 64, 128, 256, and finally 512 in the bottom layer. This expansion strategy of feature channels enables the network to extract the macroscopic topology, sphericity deviation, and overall volumetric morphology features of microplastics at deeper layers. The bottleneck layer at the bottom of the network aggregates high-dimensional semantic features that have undergone multiple compressions, characterizing the abstract chemical property categories of microplastic particles.

[0076] The decoder and encoder are symmetrically distributed. The decoder upsamples deep feature maps through 3D transposed convolution operations, gradually restoring the 3D spatial dimensions of the data. At each level of the decoder, the system uses a skip connection mechanism to concatenate the feature map of the corresponding encoder level with the currently upsampled feature map along the channel dimension. This physical connection structure directly transmits the shallow high-resolution detail information captured by the encoder to the decoder, supplementing the spatial edge information lost due to downsampling. The convolutional layers in the decoder then reprocess the concatenated mixed features, effectively fusing deep semantic category information with shallow spatial boundary information, thereby generating a segmentation mask with clear edges and well-defined categories. The network's output layer maps high-dimensional features to specific chemical component category probability maps through a convolution operation with a kernel size of 1, 1, 1, achieving voxel-by-voxel classification of microplastic components.

[0077] See attached document Figure 2 This invention provides a hierarchical label design strategy, which is integrated into the supervised training process at the network output. This strategy aims to decompose the complex task of identifying microplastic chemical components into multiple sub-tasks with a logically progressive relationship. Instead of directly forcing the network to distinguish all subdivided polymer types in the initial stage, the system constructs a three-level pyramid-shaped labeling system that includes background separation, group classification, and specific identification.

[0078] At the first level of the labeling hierarchy, the system defines the task of binarizing foreground and background segmentation. The network first learns to identify whether a voxel belongs to the microplastic particle itself or to the surrounding filter membrane pores, water medium, or air background. This level establishes the geometric boundaries of the microplastic particle in three-dimensional space, generating a basic binary mask that defines the spatial region of interest for subsequent chemical composition analysis. At the second level of the labeling hierarchy, the system classifies polymers based on the similarity of their chemical structures. Specifically, the system further subdivides the microplastic foreground into polyolefins and non-polyolefins. Since polyolefin materials such as polyethylene and polypropylene have high similarity in infrared spectral characteristics and physical density, this intermediate level of classification allows the network to prioritize learning the common morphological and textural features of these two categories, reducing the risk of feature confusion from direct subcategorization.

[0079] At the third level of the labeling hierarchy, the system performs specific polymer type identification. At this stage, the network outputs specific chemical component category labels, including but not limited to polyethylene, polypropylene, polystyrene, polyvinyl chloride, and polyethylene terephthalate. To effectively handle the multi-category output at this level and the presence of copolymers or composite plastics in real-world samples, the system employs one-hot encoding technology to process the voxel labels. One-hot encoding converts the category label, originally represented by a single integer, into a multi-dimensional vector, the length of which is equal to the total number of component categories to be identified. In the vector, the index position corresponding to the actual chemical component of the voxel is set to 1, and the remaining positions are set to 0.

[0080] This one-hot encoding mechanism enables the network to handle complex situations involving the coexistence of multiple components. For composite microplastic particles formed by the physical fusion of two or more polymers, their corresponding voxel labels can simultaneously exhibit high values ​​at multiple index positions in the one-hot encoding vector. During training, this encoding method allows the loss function to independently calculate the prediction error for each category, thereby driving the network to learn the feature responses of the corresponding chemical components on different channels of the output layer. Through this hierarchical design combined with one-hot encoding, the model can gradually transition from macroscopic object presence detection to microscopic molecular type identification, achieving a precise mapping from morphological segmentation to chemical recognition.

[0081] See attached document Figure 2 This invention provides a loss function calculation and parameter optimization strategy, configured to guide the iterative update of deep neural network weights to maximize the accuracy of microplastic component identification. Considering that microplastic samples often occupy only a very small proportion of the imaging volume, while background areas such as filter membrane pores or water media occupy the majority of the volume, this significant foreground-background imbalance causes the traditional pixel-wise cross-entropy loss function to overemphasize background classification during training, thus neglecting the segmentation accuracy of small particles. Therefore, this invention uses the Dice loss function based on overlap measurement as the objective function. This loss function directly calculates the similarity coefficient between the predicted volume and the true label volume, achieving high-sensitivity segmentation of small targets by maximizing their intersection and minimizing their difference. Before calculating the loss, the system first applies a Sigmoid activation function to the feature map of the network output layer, mapping the output value of each voxel to a probability value between zero and one, representing the confidence that the voxel belongs to a specific chemical component category.

[0082] The formula for calculating the Dice loss function is as follows:

[0083] ;

[0084] In the above formula, This represents the calculated Dice loss value. The closer the value is to zero, the better the prediction matches the actual situation. This indicates the total number of voxels contained in the input 3D volume data. Indicates the first The probability value of an individual element being predicted by the network as the target chemical component. Indicates the first The true label value of a voxel is one if the voxel actually belongs to the target chemical component, and zero otherwise. It is a very small smoothing constant used to prevent the denominator from being zero and to maintain the stability of numerical calculations. This formula strengthens the reward for the overlapping part of the predicted region and the real region through the summation of the product of the numerator, enabling the network to effectively resist the gradient vanishing problem caused by the dominance of background voxels during backpropagation, thereby establishing the ability to accurately segment the edges and internal textures of small particles.

[0085] In the optimization of network parameters, this invention employs the Adam optimizer, an adaptive moment estimation algorithm. This optimizer combines the momentum of gradients and the root mean square propagation characteristics, dynamically adjusting the learning rate based on the first and second moment estimates of each parameter, thereby achieving efficient convergence on sparse gradient 3D voxel data. To further find the global optimum and prevent training from getting stuck in local minima, the system integrates a dynamic learning rate decay strategy based on validation set performance metrics. Specifically, the system monitors the component recognition accuracy on the validation set in real time during training. When this accuracy does not show a significant improvement within a preset number of training epochs, the system automatically triggers the learning rate decay mechanism, multiplying the current learning rate by a decay factor less than one. This strategy allows the network to quickly approach the optimal solution region with a larger step size in the early stages of training, while finely adjusting the weights with a smaller step size in the later stages, thereby improving the model's ability to fit the complex aging texture features of microplastics.

[0086] See attached document Figure 1 This invention provides a microplastic composition inference detection workflow, configured to convert the morphological data of a test sample into a three-dimensional segmentation result with chemical composition semantics. In the initial step of the inference phase, an imaging device performs a non-destructive scan of the microplastic test sample with unknown chemical composition. This scanning process only acquires morphological image sequences containing information on the sample's physical morphology, internal texture, and surface roughness, without acquiring any spectrochemical information. The raw data output by the imaging device is typically a two-dimensional slice sequence in a digital imaging and communication standard format or a labeled image file format.

[0087] The computing system receives the aforementioned raw two-dimensional slice sequence and immediately initiates a data preprocessing procedure. This procedure performs data normalization operations identical to those used in the model training phase to ensure that the retained feature distribution matches the model weights. The system first reads the spatial localization parameters from the sequence metadata and uses an image processing library to stack the discrete two-dimensional slices along the Z-axis, reconstructing a continuous three-dimensional voxel matrix. Subsequently, the system converts this three-dimensional matrix into a Neuroimaging Informatics Technology Initiative (NITI A) format with a standard physical coordinate system definition. During this process, the system normalizes the voxel grayscale values ​​and resamples the volumetric data according to preset spatial resolution parameters to ensure that the voxel spacing of the test samples remains strictly consistent with the voxel spacing of the training dataset.

[0088] After the data is ready, the system loads the pre-trained DynUNet deep neural network model parameters from the storage medium. The system puts the model into inference mode, freezing the weights of all convolutional and normalized layers to prevent gradient updates during inference. Preprocessed 3D microplastic volume data is input as a tensor into the network model. The encoder path of the network model extracts low-level geometric features and high-level texture features from the input data layer by layer, while the decoder path maps these features back to the original spatial resolution. At the network output, the model does not directly generate class labels but instead outputs a multi-channel probabilistic feature map, where each channel corresponds to a predefined chemical composition category or background category.

[0089] To obtain the final component identification result, the system performs a maximum value indexing operation on the output multi-channel probabilistic feature map. For each voxel in the volume data, the system compares its predicted values ​​across all category channels and uses the channel index with the largest value as the predicted category label for that voxel. This operation generates a discretized 3D segmentation mask with the exact same spatial dimensions as the input data. Each integer value in this mask corresponds to a specific microplastic polymer type or environmental background, thus achieving the technical effect of inferring the chemical composition of microplastics on a voxel-by-voxel basis in 3D space based solely on morphological input.

[0090] See attached document Figure 1 This invention provides a reasoning detection process, which corresponds to Figure 1Step S5 aims to automate the component identification and segmentation of the microplastic sample using a trained deep neural network model. In the initial stage of inference, the user scans the environmental filter sample using a computed tomography (CT) microscope or an optical microscope with Z-axis scanning capability, acquiring a series of two-dimensional slice images containing the morphological characteristics of the microplastic particles. These raw images are typically stored in medical digital imaging communication formats or labeled image file formats, containing only grayscale information reflecting the sample's physical density or optical transparency, and do not yet contain any chemical composition information.

[0091] Upon receiving the original slice sequence, the system immediately initiates an automatic preprocessing procedure. This procedure strictly reuses the data standardization logic from the training phase. First, it iterates through the image file's metadata, reading the unique identifiers of sequence instances to identify and group slice sets belonging to the same physical sample. Then, the system calls an image processing tool library to reconstruct the two-dimensional slice sequence into three-dimensional volume data, reading the pixel spacing and origin coordinates from the metadata. The system converts the reconstructed data into the standard NIfTI format, establishing the absolute position and size of voxels in physical space, thereby generating standardized test volume data that perfectly matches the dimensions of the model's input layer. This process ensures that the distribution of input data during inference remains consistent with the distribution of training data, eliminating recognition errors caused by differences in device parameters.

[0092] After the data is ready, the system loads the optimal model weights selected from the validation set during the training phase. These parameters are loaded into the pre-built DynUNet network architecture. The system inputs the standardized test volume data into the network and performs forward propagation computation. During this process, the network uses deep convolutional kernels to extract the macroscopic geometric features and microscopic surface texture features of the test samples, and restores these features to the original spatial resolution step by step in the decoder through skip connections. The network does not perform backpropagation or parameter updates during the inference phase, but instead keeps the weights frozen to ensure the determinism and reproducibility of the output results.

[0093] The network's output layer generates a multi-channel probability volumetric map, where each channel corresponds to a chemical composition category in a predefined label hierarchy. The system compares the output values ​​of each voxel location across all channels and selects the channel index with the highest probability value as the final classification label for that voxel. Through this operation, the system outputs a 3D segmentation mask file. In this mask file, the background region is marked as zero, while each voxel identified as a microplastic is assigned an integer code corresponding to a specific polymer type, thus completely reconstructing the chemical composition distribution model of the microplastic particles in 3D space.

[0094] See attached document Figure 1This invention provides a method for result post-processing and report generation, configured to convert discrete 3D segmentation masks output by a neural network into environmental monitoring data with physical metric significance. After obtaining the segmentation mask containing chemical composition encoding, the system first reads the spatial metadata stored in the NIfTI header file and extracts the physical spacing parameters of voxels in three orthogonal directions. The system determines the minute physical volume represented by a single voxel by calculating the product of these three spacing parameters. Subsequently, the system traverses the entire 3D mask array and, for each preset polymer category, counts the total number of voxels labeled as that category.

[0095] The system uses the following volume quantization formula to accurately calculate the first... The absolute physical volume of microplastic-like components (e.g., polyethylene) in the sample being tested. :

[0096] ;

[0097] In the above formula, This represents the total number of voxels predicted as class k in the 3D segmentation mask (i.e., the statistical mask value). (number of voxels); , , These represent the physical voxel spacing in three orthogonal directions extracted from the NIfTI file header information, typically in micrometers. This formula transforms discrete computer vision counts into physically meaningful volume measures, thereby accurately calculating the absolute volume of polyethylene, polypropylene, polystyrene, and other microplastic components in the sample being tested.

[0098] The system further sums up the volumes of all detected microplastic components to obtain the total microplastic volume of the sample, and calculates the percentage of each individual component's volume in the total volume, using this as a key indicator to assess the specific plastic pollution load in water or soil samples.

[0099] To extract individual particle information from continuous voxel stacks, the system performs 3D connected component analysis on the segmentation mask. The system uses the 26-neighborhood connectivity rule to search in 3D space, merging spatially adjacent sets of voxels with the same chemical label into independent connected domains, each representing an independent microplastic particle. The system counts the number of connected domains under each chemical composition category, thus obtaining the particle count for each type of microplastic. For each identified individual particle, the system calculates its minimum bounding rectangle or principal axis direction, extracts the maximum Feret diameter as a length index, extracts the minimum Feret diameter as a width index, and calculates the length-to-width ratio. Using this aspect ratio, the system automatically classifies particles into fibrous, fragmented, or spherical shapes, thereby revealing the morphological distribution characteristics of microplastics.

[0100] Finally, the system integrates the calculated physical volume distribution, particle number statistics, and morphological characteristics into a structured format, automatically generating an environmental monitoring report. This report clearly lists the sample's unique identifier, the total abundance of microplastics, the volume percentage of each chemical component, and the particle concentration distribution across different particle size ranges. The system renders this report as a visual chart or exports it as a standard data file, allowing environmental scientists to use it directly for pollution source tracing analysis or ecological risk assessment. This achieves fully automated detection from low-cost morphological imaging to high-precision quantitative analysis of chemical components.

[0101] Specific application examples:

[0102] To verify the application effect of the multimodal data-driven rapid identification method for microplastic components of this invention in actual environmental monitoring, this embodiment selects the surface water of a river estuary as the monitoring object. After sampling, the sample containing microplastic particles was enriched on a glass fiber filter membrane with a diameter of 47 mm after standard digestion and flotation pretreatment.

[0103] In the data acquisition and reconstruction stage of step S1, the filter membrane sample was scanned using a micro-CT device with a resolution of 5 μm. The system first reads the metadata of the original DICOM sequence, identifies the unique identifier of the sequence instance, and groups 600 two-dimensional slices belonging to that sample. Subsequently, the system performs three-dimensional reconstruction, extracting the voxel spacing parameters from the metadata, including the X-axis spacing. Y-axis spacing μm, Z-axis interlayer spacing μm. The system converts the reconstructed 3D matrix into NIfTI format, at which point the data has been normalized to an isotropic voxel space.

[0104] See attached document Figure 3 , Figure 3This figure illustrates the standardized 3D voxel grayscale feature distribution curves extracted by the system after preprocessing batch samples in this embodiment. The horizontal axis represents the normalized voxel grayscale intensity, ranging from 0.00 to 1.00; the vertical axis represents the voxel frequency distribution density, with a maximum normalized value of 1.00. As shown, the curve exhibits a typical bimodal distribution with good separation between the two peaks. The left peak corresponds to the background region peaks such as water, air, or filter membrane pores, with its grayscale center precisely located at 0.20 and a corresponding relative frequency density of approximately 0.50. The right peak corresponds to the high-density microplastic region peaks, with its grayscale center located at 0.75, and its relative frequency density reaches a peak of 1.00 due to the concentrated target features. A clear trough exists between the two peaks, indicating that after the standardization processing of this invention, the target and background in the morphological data are statistically significantly separable, providing a high-quality input foundation for subsequent feature extraction by the network.

[0105] In the inference and environment report generation stage of step S5, the system inputs the preprocessed 3D voxel data into the trained DynUNet network. The network utilizes the aforementioned grayscale distribution features and spatial texture features to output a 3D segmentation mask. The system identifies a single microplastic particle, whose predicted category is polyethylene (category code 1). The system counts the total number of voxels occupied by this particle in the mask as NPENPE, which is 8500. Based on the volume quantization formula of the present invention... Calculation: Physical volume of a single voxel = 5μm × 5μm × 5μm = 125μm 3 The absolute physical volume of the polyethylene particle .

[0106] Experimental verification and effect comparison:

[0107] To verify the superiority of the proposed method based on spatial transformation enhancement and the DynUNet network over existing technologies, a validation dataset containing 500 microplastic samples (300 polyethylene and 200 polypropylene) was constructed. Two control groups were set up for the experiment:

[0108] Control group A (traditional method control group): Only two-dimensional optical microscope images were used, combined with a traditional support vector machine (SVM) classifier, relying on manually designed shape features for identification.

[0109] Experimental group (method of this invention): Three-dimensional morphological data were acquired using micro-CT, enhanced with forty-eight spatial dimensions, and trained end-to-end using DynUNet.

[0110] The main evaluation index of the experiment is the Dice coefficient.

[0111] See attached document Figure 4 , Figure 4 The graph details the performance convergence of the two methods during model training. The horizontal axis represents the number of training iterations, marked with four key nodes: 0, 100, 400, and 500. The vertical axis represents the Dice similarity coefficient on the validation set, marked with five key scales: 0.00, 0.65, 0.70, 0.89, and 1.00.

[0112] like Figure 4 As shown by the solid line, the method of this invention benefits from the three-dimensional spatial data augmentation strategy, resulting in extremely high model learning efficiency, with the curve showing a rapid upward trend. As can be seen from the solid data points marked in the figure, the Dice coefficient rapidly climbed to 0.70 by the 100th training round; as training continued, the slope of the curve gradually flattened and tended to saturate, stabilizing at 0.89 by the end of the 500th round.

[0113] In comparison, such as Figure 4 As shown by the dashed line, the traditional method's control group, lacking the ability to map deep spatial features to chemical composition, exhibits slow performance improvement and a low upper limit. At the end of the same 500th training round, its Dice coefficient only reaches 0.65.

[0114] Experimental results show that the method of the present invention is significantly better than the traditional method in terms of recognition accuracy (0.89 vs 0.65), and exhibits superior convergence performance (0.70) in the early stage of training (100 rounds), which fully demonstrates the effectiveness and robustness of establishing the mapping relationship between morphology and composition through multimodal data.

Claims

1. A rapid identification method for microplastic components based on multimodal data-driven methods, characterized in that, Includes the following steps: Morphological sequence images and chemical composition maps of microplastics are acquired. The chemical composition maps are parsed and mapped into three-dimensional voxel-level component labels. A standardized three-dimensional voxel dataset containing morphological features and component labels is constructed through physical space-consistent metadata alignment. The process of constructing the standardized three-dimensional voxel dataset through physical space-consistent metadata alignment includes: grouping the original two-dimensional slice sequences of the morphological sequence images using unique identifiers of sequence instances to identify slice sets belonging to the same microplastic sample; stacking the slice sets into a three-dimensional voxel array along a direction perpendicular to the image plane based on the position information recorded in the metadata; converting the three-dimensional voxel array into NIFTI format and extracting the origin coordinates, orientation matrix, and voxel spacing parameters from the header information of the morphological sequence images; parsing the chemical composition map and generating a corresponding three-dimensional label file; and forcibly writing the origin coordinates, orientation matrix, and voxel spacing parameters into the corresponding three-dimensional label file to achieve alignment of the morphological sequence images and the chemical composition map in the physical coordinate system. Based on the aforementioned 3D voxel dataset, a training set decoupled from the morphological image list and the component label list is constructed, and 3D data augmentation based on spatial transformation is performed. A component recognition network was trained using a training set augmented with spatial transformation-based 3D data to establish a mapping relationship between the 3D morphological features of microplastics and the distribution of chemical components. Morphological sequence images of the sample to be tested are acquired and input into the trained component recognition network. The mapping relationship is used to output a three-dimensional segmentation mask with chemical component encoding, thereby realizing the rapid identification of microplastic components.

2. The method for rapid identification of microplastic components based on multimodal data-driven approach according to claim 1, characterized in that, The process of constructing a training set that decouples the morphological image list from the component label list includes: Configure a keyword list containing preset feature strings, which are used to identify file attributes containing true value information of chemical composition; Traverse the data storage path and write the file path whose file name contains the preset feature string into the component label list and the file path whose file name does not contain the preset feature string into the morphological image list. Read the sample metadata attributes and stratify them based on microplastic polymer type or morphology category to ensure that the proportion of each category is consistent between the training set and the validation set. The random number generator is initialized with a fixed random seed parameter to generate a set of indexed permutation vectors; The index permutation vector is used to perform synchronous rearrangement of the morphological image list and the component label list, and the training set is divided by truncation according to a preset ratio.

3. The method for rapid identification of microplastic components based on multimodal data-driven approach according to claim 1, characterized in that, The process of performing spatial transformation-based 3D data augmentation includes: A four-dimensional affine transformation matrix is ​​constructed that defines the mapping relationship from the original voxel coordinate system to the target augmented coordinate system; Scaling factors are randomly generated along the three axial directions, and the new position of the voxel in space is calculated using the four-dimensional affine transformation matrix; For the morphological sequence image data in the training set, the gray value at the new position is calculated using a trilinear interpolation algorithm; For the label data corresponding to the chemical composition maps in the training set, the nearest neighbor interpolation algorithm is used to select the original voxel label value with the closest geometric distance as the label value of the new position.

4. The method for rapid identification of microplastic components based on multimodal data-driven approach according to claim 3, characterized in that, The process of performing spatial transformation-based 3D data augmentation also includes: Based on the geometric properties of the three-dimensional Cartesian coordinate system, calculate all possible arrangements of the three spatial axes and the positive and negative reading states of each axis; By combining six axial sequences with eight directional states, forty-eight unique spatial dimension arrangement key values ​​are generated. Randomly select a spatial dimension arrangement key value, and perform dimension transpose operation and inversion operation on the samples in the training set; The dimension transpose operation and the inversion operation are applied simultaneously to the morphological input image volume and the chemical composition label volume.

5. The method for rapid identification of microplastic components based on multimodal data-driven approach according to claim 1, characterized in that, The process of training the component recognition network using the enhanced training set includes: A dynamic U-Net network architecture, comprising an encoder, a decoder, and skip connections, is constructed as the component recognition network. The geometric and texture features of microplastics are extracted using the multi-layer convolution module in the encoder, and the spatial resolution of the feature map is gradually reduced by setting the convolution stride. The decoder upsamples the deep feature map through a 3D transposed convolution operation, and then splices the corresponding feature map output by the encoder through the skip connection. The high-dimensional features are mapped to the corresponding chemical composition category probability maps by using the convolution operation of the output layer, and the parameters are updated by using the DiceLoss loss function combined with the Adam optimizer.

6. The method for rapid identification of microplastic components based on multimodal data-driven approach according to claim 5, characterized in that, The process of updating parameters using the DiceLoss loss function in conjunction with the Adam optimizer includes: The Sigmoid activation function is applied to the feature map of the output layer of the component recognition network to obtain the probability value of voxels belonging to a preset chemical component category; Calculate the intersection between the predicted volume and the true label volume, and calculate the Dice loss value based on the ratio of the intersection to the sum of the predicted volume and the true label volume; The Adam optimizer is used to dynamically adjust the learning rate based on the first and second moment estimates of the parameters; The component identification accuracy on the monitoring validation set is triggered. If the component identification accuracy does not improve within a preset period, a learning rate decay mechanism is activated.

7. The method for rapid identification of microplastic components based on multimodal data-driven approach according to claim 1, characterized in that, Establishing the mapping relationship between the three-dimensional morphological characteristics of microplastics and the distribution of chemical components includes setting up a hierarchical labeling system. The specific construction steps of the labeling system are as follows: Define a first-level label, which is used to distinguish the microplastic foreground from the environmental background; Define a second-level label to classify microplastic foregrounds into polyolefins and non-polyolefins based on chemical structural similarity; A third-level label is defined, which uses unique thermal coding technology to identify specific polymer types, including polyethylene, polypropylene, and polystyrene. The component recognition network is driven to output the predicted probabilities of the first-level label, the second-level label, and the third-level label at the output end.

8. The method for rapid identification of microplastic components based on multimodal data-driven approach according to claim 1, characterized in that, The specific steps for outputting the 3D segmentation mask with chemical composition encoding are as follows: The morphological sequence images of the test sample are preprocessed into a test three-dimensional voxel matrix consistent with the format of the standardized three-dimensional voxel dataset; The preprocessed three-dimensional voxel matrix to be tested is input into the component recognition network to obtain multi-channel probability feature maps; Perform a maximum value indexing operation on the channel dimension of the probabilistic feature map, and use the channel index with the maximum probability value as the predicted class label of the voxel; A three-dimensional segmentation mask with chemical composition encoding is generated, where the background area is marked as zero and the microplastic area is marked as an integer encoding corresponding to the polymer type.

9. The method for rapid identification of microplastic components based on multimodal data-driven approach according to claim 8, characterized in that, It also includes generating an environmental monitoring report based on the three-dimensional segmentation mask with chemical composition encoding, the specific process of which is as follows: Extract the physical spacing parameters of the three-dimensional voxel matrix to be measured in three orthogonal directions, and calculate the physical volume of a single voxel; The total number of voxels under each polymer category was counted, and the absolute volume and percentage of each microplastic component were calculated. Connectivity component analysis was performed on the three-dimensional segmentation mask with chemical composition encoding using the 26-neighborhood connectivity rule to identify individual microplastic particles; Calculate the maximum and minimum Feretta diameters of the individual microplastic particles, and classify the individual microplastic particles into fibrous, fragmented, or spherical shapes based on the aspect ratio. The absolute volume and proportion of each microplastic component are summarized with the classification results of the individual microplastic particles to generate the environmental monitoring report.

Images