An efficient cell classification method and application based on 3D refractive index point cloud and deep learning of holographic tomography

Abstract

The present invention belongs to the field of optical imaging and biomedical analysis technology, and specifically relates to a three-dimensional refractive index point cloud based on holographic tomography and deep learning. The method comprises the following steps: Step 1: Acquisition and preprocessing of three-dimensional refractive index data; Step 2: Point cloud data generation and partition balanced sampling; Step 3: Constructing an RPNet++ classification network model based on PNet++, and proposing the refractive index (RI) point cloud of the present invention on the original PointNet++ network, named RI-PointNet++ (RPNet++), which is specifically used to extract features from the cell RI point cloud; Step 4: Training the RPNet++ classification network model; Step 5: Testing the RPNet++ classification network model, and using the successfully tested RPNet++ classification network model to complete cell classification based on the cell point cloud data obtained after preprocessing. The present invention can significantly reduce the computational burden and time cost when processing large-scale cell data.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical imaging and biomedical analysis, and specifically relates to an efficient cell classification method and application based on three-dimensional refractive index point cloud and deep learning of holographic tomography. Background Art

[0002] Accurate cell classification and identification are of great significance in biology, medicine, and related fields. Efficient and accurate cell classification not only helps reveal cellular functions, behaviors, and interactions, but also facilitates drug development. However, the key to cell classification lies in obtaining high-quality imaging data and effectively extracting classification features from complex data.

[0003] Fluorescence microscopy (FM) has become a widely used technique in cell research because it labels specific cellular components and provides molecular-specific information. However, FM also has significant drawbacks, such as photobleaching and phototoxicity, which limit its application in long-term observation of living cells. Although label-free microscopy techniques (such as brightfield, darkfield, and phase contrast microscopy) avoid the interference of dye labeling on cell status, these techniques generally only provide two-dimensional qualitative data and cannot fully reveal the three-dimensional structural characteristics of cells, thereby limiting the accuracy and depth of cell classification.

[0004] Three-dimensional refractive index imaging techniques (such as holographic tomography, HT) reconstruct the three-dimensional refractive index distribution of cells by collecting phase information from multiple angles, providing accurate three-dimensional structural maps for cell classification. However, while three-dimensional refractive index imaging can provide accurate three-dimensional structural maps of cells, the high-dimensional nature of voxel data leads to extremely high computational complexity, making it difficult to meet the needs of large-scale cell classification. In addition, existing methods often rely on manually designed features or traditional machine learning algorithms for feature extraction and classification tasks, which cannot fully exploit the spatial characteristics of three-dimensional data, and the classification accuracy is limited.

[0005] With the recent development of deep learning, more and more researchers have begun to explore combining deep learning with 3D imaging data to further improve classification accuracy and efficiency. However, these methods rely on large amounts of high-dimensional data and powerful computing resources, which results in significant computational pressure and hardware requirements in practical applications. This high resource demand, especially in scenarios involving large datasets and real-time processing, continues to limit their widespread application.

[0006] In order to reduce computational complexity and improve classification efficiency, point cloud technology has gradually become a research hotspot. As an efficient three-dimensional data representation, point cloud shows significant advantages in data volume and classification efficiency. However, how to extract complex cell structure features from point cloud data, optimize point cloud distribution, and reduce the hardware requirements of deep learning models remain the main challenges facing current research. To this end, the present invention proposes an efficient cell classification method and application based on three-dimensional refractive index point cloud and deep learning based on holographic tomography. Summary of the Invention

[0007] The purpose of the present invention is to provide an efficient cell classification method and application based on three-dimensional refractive index point cloud and deep learning of holographic tomography. By constructing a three-dimensional refractive index point cloud model and combining it with deep learning technology, specifically through optimized three-dimensional point cloud data processing, advanced deep learning algorithms and efficient classification models, efficient extraction and classification optimization of cell three-dimensional structural features can be achieved, which can greatly improve the accuracy and speed of cell classification, especially when processing large-scale cell data, and can significantly reduce the computational burden and time cost.

[0008] The technical solutions adopted by the present invention are as follows:

[0009] A method for efficient cell classification based on 3D refractive index point cloud and deep learning of holographic tomography comprises the following steps:

[0010] Step 1: Acquisition and preprocessing of 3D refractive index data;

[0011] Step 2: point cloud data generation and partition balanced sampling;

[0012] Step 3: Build the RPNet++ classification network model based on PNet++; RPNet++'s Chinese name is Improved Recursive Prediction Network;

[0013] Step 4: Train the RPNet++ classification network model;

[0014] Step 5: Test the RPNet++ classification network model, and complete cell classification based on the successfully tested RPNet++ classification network model using the preprocessed cell point cloud data.

[0015] Preferably, step 1 includes the following specific steps:

[0016] Step 101: Using holographic tomography (HT) technology to obtain three-dimensional refractive index data of cells. HT technology collects phase information of cell samples from multiple angles and combines it with a reconstruction algorithm to generate three-dimensional refractive index distribution data;

[0017] Step 102: De-noising the raw refractive index data using the ODT-SAS algorithm to complete preprocessing of the three-dimensional refractive index distribution data of the cells. The full name of the ODT-SAS algorithm is Optical Diffraction Tomography - Spectral Artifact Suppression algorithm. The Chinese name of the ODT-SAS algorithm is Optical Diffraction Tomography - Spectral Artifact Suppression algorithm.

[0018] Preferably, in step 2: the three-dimensional refractive index voxel data is converted into point cloud data; the point cloud data represents the three-dimensional structure inside the cell in the form of a set of discrete points, each point containing spatial coordinates (x, y, z) and a refractive index value; significant areas inside the cell are detected based on changes in the refractive index value, and the significant areas are preferentially converted into a point cloud format to retain key three-dimensional geometric and physical properties; the internal area of ​​the cell is partitioned according to the refractive index value, and different sampling densities are used for different refractive index areas, and finally the three-dimensional refractive index point cloud data of the cell is obtained.

[0019] Preferably, in step 3, after the three-dimensional refractive index point cloud data of the cell is generated, an RPNet++ classification network model based on PNet++ is constructed, and the RPNet++ classification network model includes a refractive index point cloud selector (RIPCS) for optimizing the point cloud distribution and enhancing key area features; a multi-scale feature abstraction module for extracting local and global features of the point cloud layer by layer; and a classification module for combining features and outputting the classification results of the cell.

[0020] Preferably, the RPNet++ classification network model is specifically optimized for the characteristics of the three-dimensional refractive index point cloud data of cells, and combines the farthest point sampling (FPS) and refractive index interval-enhanced sampling (RIIES) strategies to further optimize the point cloud distribution and enhance the feature expression of significant areas, such as the feature expression of cell nucleoli and lipid droplets; gradually extracts local and global features of the point cloud through a hierarchical structure, and uses a multi-scale method to capture the complex geometric morphology and physical properties inside the cell; designs a lightweight classification layer, combines the extracted global features, and finally outputs the cell classification results.

[0021] Preferably, in step 4, after the point cloud data is generated, the training of the RPNet++ classification network model is completed by constructing and optimizing the data set; the point cloud data set is divided into a training set and a test set with a ratio of 8:2; the AdamW optimizer is used, the initial learning rate is 0.001, the weight decay factor is 0.0001, and the training cycle is set to 200 times; during the training process, the cross entropy loss function is combined with the label smoothing technology, the model parameters are updated through the back propagation algorithm, and the network is gradually optimized to improve the classification performance. The Chinese name of the AdamW optimizer is Adaptive Moment Estimation Weight Decay Optimizer.

[0022] Preferably, in step 4, random rotation, Gaussian noise, and random perturbation data enhancement strategies are added during the training process; in order to ensure the stability and robustness of the model under different data distributions, the present invention adopts the k-fold cross-validation method to evaluate and adjust the model parameters; by performing multiple training and verification, the optimal configuration of hyperparameters is ensured, thereby further improving the classification performance; in order to optimize the learning process, the present invention adopts the Cosine Annealing learning rate scheduling method to dynamically adjust the learning rate so that the model can converge stably during the training process, avoid local minimum problems, and improve the convergence effect. After multiple rounds of iterative optimization, the RPNet++ classification network model can efficiently extract key information from point cloud data, accurately distinguish different cell types, and significantly improve classification accuracy.

[0023] Preferably, in step 5, the trained RPNet++ classification network model is applied to the test data set to automatically classify the cell point cloud number obtained after preprocessing; the RPNet++ classification network model generates a classification label for each cell based on the three-dimensional spatial characteristics and refractive index information of the point cloud; the classification results are displayed through a visualization tool, and the cell type and its classification accuracy are output, and relevant statistical data are generated at the same time.

[0024] An application of an efficient cell classification method based on three-dimensional refractive index point cloud of holographic tomography and deep learning, wherein the efficient cell classification method is applied to cell classification, cell activity assessment and drug screening.

[0025] The technical effects achieved by the present invention are:

[0026] 1. This method uses three-dimensional refractive index point cloud data as a representation of the internal structure of cells. Compared to traditional two-dimensional image representations, this method can fully reflect the internal geometric structure and refractive index distribution characteristics of cells. Compared with voxel data representation, point cloud data significantly reduces the data volume while preserving key structural features, thereby significantly reducing computational complexity and improving classification accuracy and efficiency.

[0027] 2. This paper proposes a balanced sampling method that combines Refractive Index Enhanced Sampling (RIIES) with Farthest Point Sampling (FPS). This method effectively addresses the problem of traditional sampling methods underrepresenting features in key intracellular regions (such as the nucleus and nucleolus). By uniformly covering regions with varying refractive index ranges, this method significantly improves the representation of complex internal cellular structures while reducing the risk of uneven data distribution affecting classification results, ensuring greater reliability.

[0028] 3. Traditional cell classification methods typically rely on manually designed geometric or physical features and are susceptible to limitations caused by human experience and subjective judgment. However, this invention, through the improved RPNet++ classification network model, can automatically extract multi-scale local and global features from point cloud data, fully capturing the complexity of 3D cell structures, avoiding errors caused by manual feature selection, and significantly improving classification performance.

[0029] 4. The RPNet++ classification network model of this invention achieved a classification accuracy of 93.5% in the HeLa cell classification task, significantly exceeding the traditional two-dimensional method (87.0%) and the three-dimensional voxel method (82.6%). Furthermore, the computational complexity of this method was significantly reduced, with floating-point operations (FLOPs) of only 1.49 GB, a reduction of over 99% compared to traditional three-dimensional voxel models. Thanks to its lightweight design, this method can run efficiently on standard CPU devices, achieving classification performance nearly on par with GPU environments, demonstrating broad application potential.

[0030] 5. This invention is suitable for the efficient classification of large-scale cell data and can be widely used in biomedical fields such as cell activity assessment and drug screening, providing accurate and reliable technical support for exploring cell behavior and functional status. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a schematic diagram of the holographic tomography optical system of the present invention;

[0032] Figure 2 It is a visualization diagram comparing the three-dimensional refractive index voxel data and point cloud data in the present invention, where Figure 2 (ac) show the 3D refractive index voxel distribution of high-activity, medium-activity, and low-activity HeLa cells, respectively, reflecting the refractive index variation characteristics of the internal structure of the cells; Figure 2 (df) is the corresponding three-dimensional refractive index point cloud data. The point cloud format significantly reduces the amount of data while retaining the geometric and physical information of key feature areas (such as cell nuclei and lipid droplets);

[0033] Figure 3 This is the process for generating cell refractive index point cloud data and sampling strategy in the present invention, including the original three-dimensional voxel refractive index image ( Figure 3(a)), random sampling strategy image ( Figure 3 (b)) and the impact of the number of partitions (L value) on sampling effect and classification performance ( Figure 3 (c) and Figure 3 (d)); Figure 3 (eh) is the corresponding refractive index distribution;

[0034] Figure 4 This is the structural framework diagram of the RPNet++ classification network in the present invention;

[0035] Figure 5 This is the evaluation and experimental results of the classification performance in the present invention, including Figure 5 (a) shows the comparison results of classification accuracy of the six groups of experiments in Table 3; Figure 5 (b) Comparison of the F1 score, recall rate, and precision of the classification results; Figure 5 (c) shows the effect of the number of refractive index intervals (L value) in partition balanced sampling on the classification performance; Figure 5 (d) ROC curve and AUC value of the classification model, which verifies the robustness of the model of the present invention; Figure 5 (e) is the confusion matrix of the classification results, showing the classification accuracy of high, medium, and low activity cells. DETAILED DESCRIPTION

[0036] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is only used to describe one or more specific embodiments of the present invention and does not strictly limit the scope of protection of the present invention. Example

[0037] like Figure 1 As shown, a method for efficient cell classification based on 3D refractive index point cloud of holographic tomography and deep learning includes the following steps:

[0038] Step 1: Acquisition and preprocessing of 3D refractive index data;

[0039] The step 1 includes the following specific steps:

[0040] Step 101: Use holographic tomography (HT) technology to obtain three-dimensional refractive index data of cells. HT technology collects phase information of cell samples from multiple angles and combines it with a reconstruction algorithm to generate three-dimensional refractive index distribution data. This data can accurately reflect the spatial structure and physical properties of the cell interior.

[0041] Step 102: To improve data quality and reduce artifacts and noise, the ODT-SAS algorithm (Optical Diffraction Tomography - Spectral Artifact Suppression) is used to denoise the raw refractive index data. This completes preprocessing of the 3D refractive index distribution of the cells and effectively preserves the refractive index distribution of the cells. This step generates high-quality refractive index data that provides reliable input for subsequent point cloud conversion and classification tasks.

[0042] In this embodiment, Figure 1 As shown, Figure 1 This is a schematic diagram of the holographic tomography optical system, showing the experimental setup and optical component arrangement used to collect 3D refractive index data in this invention, including the beam splitter, focusing objective lens, imaging objective lens, and galvanometer mirror. This optical system reconstructs 3D refractive index data of cells by collecting multi-angle phase information.

[0043] In this example, the present invention first acquires three-dimensional refractive index data of cells using holographic tomography (HT) technology. Specifically, the experimental system employs a Mach-Zehnder interferometer optical design. Using a low-coherence light source, such as one with a central wavelength of 641 nm, the beam is controlled by a galvanometer to illuminate the cell sample at 51 different angles, covering a numerical aperture range of 0.65. The resulting multi-angle holograms are then reconstructed in three dimensions using a holographic tomography reconstruction algorithm written in MATLAB to generate three-dimensional refractive index voxel data. To address the missing cone problem, a non-negativity constraint algorithm is applied to the reconstruction process, resulting in a three-dimensional refractive index distribution that accurately reflects the internal structure of the cell. To improve data quality and eliminate artifacts and background noise, the ODT-SAS algorithm is used to denoise the voxel data. The processed refractive index distribution data has higher accuracy and resolution, providing high-quality input for subsequent data processing.

[0044] This embodiment Figure 1Middle: BS: beam splitter; GM: galvanometer mirror; L: lens; M: reflector; Obj: microscope objective; P: polarizer; PBS: polarizing beam splitter; TL: tubular lens; HWP: half-wave plate; QWP: quarter-wave plate. The system starts from the light source. Light passes through a filter (Filter) to adjust the intensity, a polarizer (P1) to adjust the polarization direction, and a half-wave plate (HWP1) to preliminarily modulate the polarization state. The light then passes through a 4f system consisting of lenses (L1) and (L2) for amplification and collimation before entering the core control region of the optical path. Within this region, polarizing beam splitters (PBS1 and PBS2) split the beam, while quarter-wave plates (QWP1, QWP2, and QWP3) adjust the polarization phase. Mirrors (M1, M2, M3, and M4) redirect the optical path. The galvanometer mirror (GM) provides dynamic beam steering, enabling rapid optical path switching or scanning. The reference light is polarized (P2) and then collimated or amplified by lenses (L3 and L4). After the object light is scanned by the galvanometer, it is also collimated or amplified by tubular lenses (TL1 and TL2). The light is focused onto the sample through the microscope objective (Obj1) and then amplified by the microscope objective (Obj2). Finally, the object light and the reference light are interfered and synthesized by the beam splitter (BS) and the polarizer (P3) to enter the camera to achieve high-resolution imaging.

[0045] Step 2: point cloud data generation and partition balanced sampling;

[0046] In step 2, to reduce data complexity and improve classification efficiency, the present invention converts three-dimensional refractive index voxel data into point cloud data; the point cloud data represents the three-dimensional structure inside the cell in the form of a set of discrete points, each point containing spatial coordinates (x, y, z) and a refractive index value; significant areas inside the cell are detected based on changes in the refractive index value, and significant areas include the cell nucleus, nucleolus, and lipid droplets, and the significant areas are preferentially converted into a point cloud format to preserve key three-dimensional geometric and physical properties; the cell interior area is partitioned according to the refractive index value, and different sampling densities are used for different refractive index areas, ultimately obtaining three-dimensional refractive index point cloud data of the cell. For example, in high refractive index areas, such as the cell nucleus, the sampling is denser to enhance the ability to express key regional features; in low refractive index areas, such as the cytoplasm, the number of sampling points is reduced to optimize computational efficiency.

[0047] In the present invention, Figure 2 As shown, Figure 2 It is a visualization of the comparison between 3D refractive index voxel data and point cloud data, where Figure 2(ac) show the 3D refractive index voxel distribution of high-activity, medium-activity, and low-activity HeLa cells, respectively, reflecting the refractive index variation characteristics of the internal structure of the cells; Figure 2 (df) is the corresponding three-dimensional refractive index point cloud data. The point cloud format significantly reduces the amount of data while retaining the geometric and physical information of significant feature areas such as cell nuclei and lipid droplets.

[0048] In the present invention, Figure 3 As shown, Figure 3 The process of generating cell refractive index point cloud data and sampling strategy in the present invention is presented, including the original three-dimensional voxel refractive index image ( Figure 3 (a)), random sampling strategy image ( Figure 3 (b)) and the impact of the number of partitions (L value) on sampling effect and classification performance ( Figure 3 (c) and Figure 3 (d)); Figure 3 (eh) is the corresponding refractive index distribution.

[0049] In practical applications, the present invention proposes a partitioned balanced sampling method for converting 3D refractive index voxel data into point cloud data, significantly reducing data dimensionality while preserving key geometric and physical features within cells. Point cloud data is represented as 3D coordinates (x, y, z) and corresponding refractive index values. First, the 3D refractive index values ​​are normalized and divided into several refractive index intervals (e.g., L = 6). Experimental validation shows that when L = 6, the point cloud achieves the most balanced coverage across different refractive index intervals, fully representing key features of high-refractive index regions (such as nucleoli and lipid droplets) while effectively avoiding oversampling of low-refractive index regions. Further increasing the number of intervals (e.g., L > 6) improves the refinement of the refractive index distribution, but the improvement in classification performance reaches a saturation point and increases data processing complexity. Furthermore, when L < 6, insufficient points are present in high-refractive index regions, resulting in reduced classification performance. Therefore, selecting L = 6 achieves the optimal balance between classification performance and computational efficiency. After processing, the original three-dimensional refractive index voxel data (approximately 4 million points) was streamlined into 10,000-point high-quality point cloud data. This not only significantly reduced the data size but also achieved a balance between the expressive power of geometric details and key features, providing optimized data support for the efficient input and training of subsequent classification models.

[0050] Step 3: Build the RPNet++ classification network model based on PNet++; propose the refractive index (RI) point cloud of the present invention on the original PointNet++ network, named RI-PointNet++ (RPNet++), which is specifically used to extract features from the cell RI point cloud;

[0051] In step 3, after the three-dimensional refractive index point cloud data of the cell is generated, an RPNet++ classification network model based on PNet++ is constructed. RPNet++ is an optimized PointNet++ network named RI-PointNet++ (RPNet++), which is specifically used to extract features from the cell RI point cloud, such as Figure 4 As shown, Figure 4 It is a structural framework diagram of the RPNet++ classification network model, which shows the main modules and processing flow of the RPNet++ classification network model designed by the present invention. The RPNet++ classification network model includes a refractive index point cloud selector (RIPCS) for optimizing point cloud distribution and enhancing key area features; a multi-scale feature abstraction module for extracting local and global features of point clouds layer by layer; and a classification module for combining features and outputting cell classification results.

[0052] To achieve efficient classification of cell point cloud data, the present invention proposes an optimized deep learning network RPNet++. The RPNet++ classification network model is specifically optimized for the characteristics of three-dimensional refractive index point cloud data of cells. It combines the farthest point sampling (FPS) and refractive index interval-enhanced sampling (RIIES) strategies to further optimize the point cloud distribution and enhance the feature expression of significant regions, such as cell nucleoli and lipid droplets. It gradually extracts local and global features of the point cloud through a hierarchical structure and uses a multi-scale approach to capture the complex geometric morphology and physical properties inside the cell. A lightweight classification layer is designed and combined with the extracted global features to ultimately output the cell classification results.

[0053] In practical application, this invention uses a classification model designed based on PNet++: This invention proposes a deep learning classification model specifically for 3D refractive index point cloud data—the RPNet++ classification network model. This RPNet++ classification network model consists of three components: a Refractive Index Point Cloud Selector (RIPCS), a feature abstraction module, and a classification module. It aims to efficiently extract complex geometric and physical features from point cloud data and output accurate classification results. First, the RIPCS module optimizes and filters the input point cloud data. RIPCS combines the Farthest Point Sampling (FPS) and Refractive Index Interval Enhanced Sampling (RIIES) methods to select a representative set of key points from the input of 10,000 points and prune it to 1,024 points. In this process, the FPS method ensures uniform global coverage, while RIIES enhances features in key regions such as the nucleolus, ensuring representativeness and diversity of the sampled points. Furthermore, RIPCS introduces a randomized sampling strategy to improve the model's generalization to diverse data distributions. Next, the feature abstraction module performs hierarchical feature extraction from the point cloud using a three-layer nested Set Abstraction (SA) structure. Each SA layer comprises three steps: sampling, grouping, and RPNet feature extraction. It utilizes a ball query method to generate local region point sets. A multi-layer perceptron (MLP) and max pooling operation are then used to gradually abstract the local-to-global feature representation. The RPNet layer incorporates local coordinate transformations to eliminate interference from global position information, more accurately capturing the region's geometric and physical features. This module effectively extracts multi-scale information from the point cloud data and enhances the representation of complex cellular structures. Finally, the classification module, consisting of three fully connected layers (FC1, FC2, and FC3), progressively compresses feature dimensions and outputs a classification probability distribution for each cell. Dropout regularization is applied to the first (FC1) and second (FC2) layers, with dropout ratios of 0.3 and 0.4, respectively, to mitigate overfitting. ReLU activation is used to enhance feature representation. The final layer (FC3) maps features to the target classification space using a log-softmax activation function and outputs the classification result for each cell. Through the RPNet++ classification network model, the present invention can fully capture the key features of three-dimensional refractive index point cloud data and achieve efficient and accurate cell classification.

[0054] Step 4: Train the RPNet++ classification network model;

[0055] In step 4, after the point cloud data is generated, the RPNet++ classification network model is trained by constructing and optimizing the data set; the point cloud data set is divided into a training set and a test set with a ratio of 8:2; the AdamW optimizer is used, the initial learning rate is 0.001, the weight decay factor is 0.0001, and the training cycle is set to 200 times; during the training process, the cross entropy loss function is combined with the label smoothing technology, the model parameters are updated through the back propagation algorithm, and the network is gradually optimized to improve the classification performance.

[0056] In step 4, in order to enhance the generalization ability of the model and prevent overfitting, the present invention adds random rotation, Gaussian noise and random perturbation data enhancement strategies during the training process; in order to ensure the stability and robustness of the model under different data distributions, the present invention uses the k-fold cross-validation method to evaluate and adjust the model; by performing multiple training and verification, the optimal configuration of hyperparameters is ensured, thereby further improving the classification performance; in order to optimize the learning process, the present invention adopts the Cosine Annealing learning rate scheduling method to dynamically adjust the learning rate so that the model can converge stably during the training process, avoid local minimum problems, and improve the convergence effect. After multiple rounds of iterative optimization, the RPNet++ classification network model can efficiently extract key information from point cloud data, accurately distinguish different cell types, and significantly improve classification accuracy. The present invention provides biological researchers with in-depth analysis of cell functional status through efficient and accurate classification results, thereby assisting in drug screening tasks.

[0057] In practical application, the processed point cloud data (10,000 points) was further reduced to 1,024 points using the Refractive Index Point Cloud Selector (RIPCS). The data was then divided into a training set and a test set with a ratio of 8:2, with the training set containing 364 samples and the test set containing 92 samples. The effectiveness of the model was verified using HeLa cells as an experimental example. The experimental data covered 456 cell samples, representing three states of high, medium, and low activity. It is important to note that this method has broad applicability, and the point cloud generation and sampling strategies can be adjusted according to the characteristics of different cell types. During training, the AdamW optimizer was used with an initial learning rate of 0.001 and a weight decay factor of 0.0001. The learning rate was dynamically adjusted using a cosine annealing schedule to ensure stable model convergence. The loss function used a cross-entropy loss combined with a label smoothing technique (α = 0.1), effectively improving the model's robustness to noisy data. In addition, to improve the generalization ability of the model, data augmentation strategies such as random rotation, random perturbation, and Gaussian noise were incorporated during training. All input data was normalized before entering the model. To evaluate the robustness and stability of the model, the RPNet++ classification network model was adjusted and validated using the k-fold cross-validation method. After training, the optimized RPNet++ classification network model was used to classify the test set. The model uses the spatial geometry information and refractive index characteristics of the point cloud to predict cell type labels. Experimental results show that the RPNet++ classification network model achieved a final classification accuracy of 93.5% in the HeLa cell activity classification task, significantly higher than traditional two-dimensional image methods (87.0%) and three-dimensional voxel methods (82.6%). In addition, due to its lightweight design, the computational complexity of the RPNet++ classification network model is significantly reduced, supporting efficient operation on ordinary CPU devices, demonstrating good hardware adaptability and practical application potential.

[0058] Figure 5 This is the evaluation and experimental results of the classification performance in the present invention, including Figure 5 (a) shows the comparison results of classification accuracy of the six groups of experiments in Table 3; Figure 5 (b) Comparison of the F1 score, recall rate, and precision of the classification results; Figure 5 (c) shows the effect of the number of refractive index intervals (L value) in partition balanced sampling on the classification performance; Figure 5 (d) ROC curve and AUC value of the classification model, which verifies the robustness of the model of the present invention; Figure 5 (e) is the confusion matrix of the classification results, showing the classification accuracy of high, medium, and low activity cells.

[0059]

[0060] Table 1 shows the specific configuration parameters of the feature abstraction module. This module consists of three SetAbstraction layers (SA1, SA2, and SA3) for layer-by-layer feature extraction. The number of sampling points, receptive field radius, and MLP parameters of each layer are optimized. The Abstraction layer is also known as the Set Abstraction Layer.

[0061] Table 1 describes the structure and parameter configuration of the point cloud feature abstraction module. The feature abstraction module (Set Abstraction) processes point cloud data in a layered manner, with each layer performing three steps: sampling, grouping, and feature extraction (Rpnet). In the first layer, 512 central points are randomly sampled, and 32 points are sampled within a radius of 0.15. Multi-layer perceptrons (MLPs) are then used to extract features, resulting in an output dimension of [64, 64, 128]. In the second layer, the number of central points is reduced to 256, the radius is increased to 0.3, and the number of sampled points per group is increased to 64, further increasing the feature dimension to [128, 128, 256]. The final layer no longer performs local sampling and grouping, but instead extracts global features from all points as a whole, expanding the feature dimension to [256, 512, 1024].

[0062]

[0063] Table 2 lists the specific parameter configurations of the classification module. This module consists of three fully connected layers, which gradually compress the feature dimensions and generate classification results.

[0064] Specifically, Table 2 describes the structure and parameter configuration of a three-layer fully connected network, mainly used for classification tasks. Layer: layer, Input Dim: input dimension, Output Dim: output dimension, Dropout Rate: dropout rate, Activation: activation function;

[0065] The first layer (Fully-connected 1) receives the 1028-dimensional global feature vector as input and reduces it to 512 dimensions. It applies the ReLU activation function to introduce nonlinear representation and uses Dropout regularization (at a rate of 0.3) to prevent overfitting. The second layer (Fully-connected 2) receives the 512-dimensional output of the first layer and further compresses the features to 256 dimensions. It also uses the ReLU activation function to enhance feature representation and increases the Dropout rate to 0.4 to further enhance the model's generalization. The final layer (Fully-connected 3) maps the 256-dimensional output of the second layer to a 3D space to represent the three categories of the classification task. This layer uses the Log-Softmax activation function to transform the output features into a logarithmic probability distribution, adapting to the requirements of multi-classification tasks and improving numerical stability.

[0066]

[0067] Table 3 shows the impact of different sampling methods on point cloud generation and cell classification performance. The table lists six sets of experiments, each using a different voxel-to-point cloud sampling method (such as random sampling and partition-balanced sampling) and a refractive index point cloud selector (RIPCS) strategy (such as farthest point sampling (FPS) or hybrid sampling).

[0068] Table 3 specifically shows the parameter combinations used in different experimental settings for point cloud processing tasks, including input dimensions, voxel-to-point cloud sampling methods, and the Refractive Index Point Cloud Selector (RIPCS) strategy. Input dimensions are divided into two categories: x, y, and z represent the three-dimensional spatial coordinates of the point cloud, while x, y, z, and n are further combined with refractive index (n) information to describe physical properties and optical behavior. This combination not only provides point location information but also enhances the representation of surface geometry and material properties. In the voxel-to-point cloud sampling method, random sampling directly selects points from the point cloud. This is simple and efficient, but may result in uneven point distribution. Segmented Equilibrium Sampling, on the other hand, divides the point cloud into segments and evenly samples within each segment, helping to improve the global uniformity of the sampled points. In the RIPCS strategy, FPS sampling (Farthest Point Sampling) ensures the spatial uniformity of sampling points by gradually selecting points that are farthest from each other; Hybrid Sampling combines the strategies of Farthest Point Sampling (FPS) and Refractive Index Enhanced Sampling (RIIES). RIIES adjusts the sampling weight according to the refractive index distribution of the points, which can more accurately extract key points containing optical characteristics.

[0069]

[0070] Table 4 is a performance comparison table of the present invention and existing methods, which quantitatively compares the classification accuracy, F1 score, FLOPs, and hardware adaptability of RPNet++ with traditional two-dimensional models (such as DenseNet121-2D) and three-dimensional voxel models (such as DenseNet121-3D), highlighting the performance advantages and hardware adaptability of the present invention.

[0071] Table 4 specifically compares the performance of multiple deep learning models in terms of input data size (Input Data Size), number of parameters (M), floating-point operations (FLOPs), and classification accuracy (Classification Accuracy). DenseNet121 is divided into two-dimensional and three-dimensional models, processing two-dimensional image data (e.g., 320×320) and three-dimensional volume data (e.g., 320×320×210), respectively. ResNet101-2D achieves 87.0% classification accuracy with 42.50M parameters and 250.88G FLOPs, while MobileNetV1-2D achieves 80.4% accuracy with 3.22M parameters and 0.40G FLOPs. While RPNet++ and RPNet++ (CPU) our networks, with input data (e.g., 1024×4) after point cloud processing, have only 1.47 M parameters and 1.49 G FLOPs, achieving the highest classification accuracy of 93.5% and 92.4%, respectively.

[0072] Step 5: Test the RPNet++ classification network model and use the successfully tested RPNet++ classification network model to complete cell classification based on the cell point cloud data obtained after preprocessing.

[0073] In step 5, the trained RPNet++ classification network model is applied to the test data set to automatically classify the cell point cloud obtained after preprocessing; the RPNet++ classification network model generates a classification label for each cell based on the three-dimensional spatial characteristics and refractive index information of the point cloud; the classification results are displayed through a visualization tool, and the cell type and its classification accuracy are output, and relevant statistical data are generated at the same time.

[0074] In practical applications, the RPNet++ classification model utilizes only 1.49 GB of computational complexity (FLOPs), a reduction of over 99% compared to the traditional 3D voxel-based approach (DenseNet121-3D, 882.56 GB), significantly reducing computational burden and hardware resource requirements. This approach supports efficient operation on standard CPUs, with classification performance (92.4% accuracy) approaching that achieved on GPUs (93.5% accuracy), demonstrating strong hardware compatibility. By combining Farthest Point Sampling (FPS) with Refractive Index Enhanced Sampling (RIIES) with a partition-balanced sampling strategy, the point cloud data comprehensively covers the geometric and physical characteristics of cells and significantly improves classification accuracy. Experimental results show that random sampling yields a classification accuracy of only 83.7%, while partition-balanced sampling increases this to 93.5%. Furthermore, thanks to its lightweight feature abstraction module and optimized classification layer design, the RPNet++ network efficiently extracts both local and global features from cell point clouds, enabling precise differentiation between high-, medium-, and low-activity cells. This method not only significantly improves the accuracy and efficiency of cell classification but also reduces the model's reliance on hardware resources. Experimental results demonstrate that this method is suitable for biomedical fields such as cell viability assessment and drug screening, providing an efficient, accurate, and reliable technical solution for cell analysis with broad potential for practical application.

[0075] The present invention Figure 2 This figure shows the three-dimensional refractive index (RI) distribution and its associated characteristics at different cell viability states, representing optical imaging results for high-viability, moderate-viability, and low-viability cells, respectively. Figures (a), (b), and (c) show the three-dimensional RI distributions of high-viability, moderate-viability, and low-viability cells, with the refractive index range marked by a color map from low (blue) to high (red). In high-viability cells (a), nucleoli and lipid droplets are highlighted by high-refractive index regions (indicated by red and yellow arrows). Moderate-viability cells (b) exhibit a decreased RI distribution, with clear nucleolar boundaries but fewer and less distinct lipid droplets. Low-viability cells (c) exhibit a more uniform low-RI distribution, with blurred nucleolar boundaries and significantly fewer or absent lipid droplets. Figures (d), (e) and (f) are the three-dimensional point cloud models of high-activity, medium-activity and low-activity cells, respectively, which show the spatial distribution characteristics of the refractive index in a more intuitive way.

[0076] The present invention Figure 3Figure (a) shows a 3D refractive index voxel map (3D RI Voxel). The color range from blue (low refractive index, 1.3342) to red (high refractive index, 1.4060) indicates the refractive index distribution. High refractive index regions correspond to high-density structures within the cell, such as the nucleolus (indicated by the yellow arrow) and lipid droplets (indicated by the red arrow). Figures (b), (c), and (d) show the point cloud results generated by random sampling (Random Sampling), segmented equilibrium sampling (L=2), and segmented equilibrium sampling (L=6), respectively. Random sampling (Figure b) results in an uneven point cloud distribution and poor coverage of high refractive index regions. Segmented equilibrium sampling significantly improves point cloud uniformity and coverage of high refractive index regions by segmenting the refractive index range and sampling evenly within each segment. Figures (e), (f), (g), and (h) are histograms of voxel counts and point counts, used to quantify the sampling effect of the refractive index distribution. Figure (e) shows the overall voxel number distribution, which gradually decreases with increasing refractive index. The random sampling in Figure (f) is significantly insufficient in the high refractive index range (>1.38). The piecewise balanced sampling (L=2) in Figure (g) increases point cloud coverage in high refractive index regions, and the piecewise balanced sampling (L=6) in Figure (h) further enhances the sampling density in high refractive index regions, making the point cloud distribution more uniform.

[0077] The present invention Figure 4Figure (a) shows the overall network structure. Point cloud data is first input. The initial point cloud contains 1000 points, each with four features (e.g., x, y, z, n). After passing through the RIPCS (Refractive Index Point Cloud Sampling) module, the point cloud is divided into two parts: the 800-point point set Sfps obtained through FPS (Farthest Point Sampling), and the 224-point point set Sri obtained through Refractive Index Interval-Enhanced Sampling (RIIES). These two parts are merged to form a new point cloud (1024 points, 4D features), which then enters the subsequent feature abstraction module. The Feature Abstraction Module consists of three levels of set abstraction (SA): SA1, SA2, and SA3. Each SA layer samples, groups, and extracts features from the point cloud, continuously abstracting the spatial and geometric features of the point cloud and gradually extracting global features. After feature abstraction is completed, the point cloud features are passed to the Classification Module, gradually reduced in dimension through three fully connected layers (FC1, FC2, and FC3) and mapped to the classification output space, ultimately generating a class score.

[0078] Figure (b) details the workflow of the Set Abstraction (SA) module. First, the Sampling Layer downsamples the point cloud using Farthest Point Sampling (FPS) to select keypoints. Next, the Grouping Layer uses a ball query method to find local neighborhood points for each sampled point and group them. Finally, the grouped point cloud enters the RpNet layer, where local features are extracted using a multi-layer perceptron (MLP) and max pooling.

[0079] The present invention Figure 5 middle, Figure 5 (a) Comparison of the classification accuracy of the six experimental groups in Table III. The results show that the experimental group using segmented balanced sampling and refractive index enhancement strategy (Experiment 6) has the best performance. Figure 5(b) shows the F1 scores of the six groups of experiments: combining precision and recall, it is an indicator to comprehensively measure the classification performance of the model, precision (Precision): the proportion of samples predicted to be a certain class that are actually correct and recall (Recall): the proportion of samples that are correctly predicted to belong to a certain class. The comparison shows that the experimental group using segmented balanced sampling and refractive index enhancement strategy (Experiment 6) has the best performance. Figure 5 (c) We further analyzed the effect of the number of refractive index intervals (L) on classification performance. As the L value increased (e.g., from 2 to 6), the classification accuracy gradually improved. Further increasing the number of intervals (e.g., L>6) improved the refinement of the refractive index distribution, but the improvement in classification performance reached saturation and increased the complexity of data processing. Therefore, choosing L=6 achieved the best balance between classification performance and computational efficiency. Figure 5 (d) The robustness of our method (Experiment 6) was verified using the ROC curve (Receiver Operating Characteristic Curve), which represents the performance of the classification model, and the AUC value (Area Under the Curve) (the closer the value is to 1, the better the classification performance). The ROC curve shows the performance of the classification model under different cell activity levels. The horizontal axis represents the I-Specificity (false positive rate), which is the proportion of negative examples incorrectly classified as positive by the model; the vertical axis represents the Sensitivity (true positive rate), which is the proportion of positive examples correctly detected by the model. The AUC value is the area under the ROC curve, ranging from 0 to 1, which measures the model's classification performance. The closer the AUC is to 1, the better the model performance. The figure shows that the AUC values ​​for each classification category (high activity, medium activity, and low activity) are close to 1, indicating that the model has strong classification robustness across different categories and excellent discriminatory ability. Figure 5 (e) shows the confusion matrix analysis. The classification accuracy of high, medium, and low activity cells are 100%, 91.67%, and 90.32%, respectively. The off-diagonal values ​​are small, indicating that the model has low classification confusion.

[0080] Example 2:

[0081] An application of an efficient cell classification method based on three-dimensional refractive index point cloud of holographic tomography and deep learning, wherein the efficient cell classification method is applied to cell classification, cell activity assessment and drug screening.

[0082] The foregoing is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. An efficient cell classification method based on 3D refractive index point cloud and deep learning based on holographic tomography, characterized by: The following steps are involved: Step 1: Acquisition and preprocessing of 3D refractive index data; Step 2: point cloud data generation and partition balanced sampling; Step 3: Build the RPNet++ classification network model based on PNet++; Step 4: Train the RPNet++ classification network model; Step 5: Testing the RPNet++ classification network model and completing cell classification based on the successfully tested RPNet++ classification network model using the preprocessed cell point cloud data; the cell classification includes: the RPNet++ classification network model generates a classification label for each cell based on the three-dimensional spatial features and refractive index information of the point cloud; The step 2 includes: The three-dimensional refractive index voxel data obtained after preprocessing is converted into point cloud data; the point cloud data represents the three-dimensional structure inside the cell in the form of a set of discrete points, each of which contains spatial coordinates (x, y, z) and a refractive index value; Detect salient areas inside cells based on changes in refractive index values ​​and preferentially convert these areas into point cloud format to preserve key 3D geometric and physical properties. The internal area of ​​the cell is partitioned according to the refractive index value, and different sampling densities are used for different refractive index areas to finally obtain the three-dimensional refractive index point cloud data of the cell; The step 3 comprises: After generating the three-dimensional refractive index point cloud data of the cells, an RPNet++ classification network model based on PNet++ is constructed. The RPNet++ classification network model consists of a refractive index point cloud selector, a multi-scale feature abstraction module, and a classification module. The refractive index point cloud selector includes farthest point sampling and refractive index interval enhanced sampling, wherein the refractive index interval enhanced sampling adjusts the sampling weight according to the refractive index distribution of the point to optimize the point cloud distribution and enhance the key area features; The multi-scale feature abstraction module is used to extract local and global features of the point cloud layer by layer to capture the complex geometric morphology and physical properties inside the cell; The classification module is used to combine local features and global features and output cell classification results.

2. The method for efficient cell classification based on 3D refractive index point cloud and deep learning based on holographic tomography according to claim 1, characterized in that: The step 1 includes the following specific steps: Step 101: Acquire three-dimensional refractive index data of cells using holographic tomography technology, wherein the holographic tomography technology collects phase information of cell samples from multiple angles and generates three-dimensional refractive index distribution data using a reconstruction algorithm; Step 102: Using the ODT-SAS algorithm, the original refractive index data is subjected to denoising processing to complete the preprocessing of the three-dimensional refractive index distribution data of the cells.

3. The method for efficient cell classification based on 3D refractive index point cloud and deep learning based on holographic tomography according to claim 2, characterized in that: In step 4, after the point cloud data is generated, the RPNet++ classification network model is trained by constructing and optimizing the data set; the point cloud data set is divided into a training set and a test set with a ratio of 8:2; the AdamW optimizer is used, the initial learning rate is 0.001, the weight decay factor is 0.0001, and the training cycle is set to 200 times; During the training process, the cross entropy loss function is combined with label smoothing technology, and the model parameters are updated through the back propagation algorithm to gradually optimize the network.

4. The method for efficient cell classification based on 3D refractive index point cloud and deep learning based on holographic tomography according to claim 3, characterized in that: In step 4, random rotation, Gaussian noise and random perturbation data enhancement strategies are added during the training process; the k-fold cross-validation method is used to evaluate and adjust the model parameters; multiple training and verification are performed, and the cosine annealing learning rate scheduling method is used to dynamically adjust the learning rate so that the model converges stably during the training process.

5. The method for efficient cell classification based on 3D refractive index point cloud and deep learning based on holographic tomography according to claim 4, characterized in that: In step 5, the trained RPNet++ classification network model is applied to the test data set to automatically classify the cell point cloud obtained after preprocessing; The classification results are displayed through visualization tools, which output cell types and their classification accuracy, and generate relevant statistics.

6. An application of an efficient cell classification method based on three-dimensional refractive index point cloud and deep learning based on holographic tomography, wherein the efficient cell classification method is the efficient cell classification method as described in any one of claims 1-5, and the efficient cell classification method is applied to cell classification, cell activity assessment and drug screening.

Images