Cytology specimen pretreatment system and method based on machine vision and intelligent decision

Abstract

The invention discloses a cytology specimen pretreatment system and method based on machine vision and intelligent decision, and relates to the technical field of biological sample treatment.The cytology specimen pretreatment system comprises an operation platform, and the operation platform is provided with a first oscillation module, a centrifugal module, a visual recognition module and a sample tube operation assembly; automatic uniform mixing and centrifugal operation of sample liquid are achieved, so that early-stage materials meeting the requirements of cytology specimens are obtained, excessive manual intervention treatment is not needed, the labor cost and the infection probability in the operation process of medical staff are reduced, the working efficiency of early-stage treatment of the cytology specimens is improved, the requirement for experience of operators is low, and the operation cost is reduced. In addition, the cytological specimen is completely prepared, the cytological wax block and the thin layer are prepared, and the supernate is reserved, so that the opportunity of grafting the cytological specimen to other detection platforms subsequently is increased, and the value of the cytological specimen in pathological diagnosis is improved.

Description

Technical Field

[0001] This invention relates to the field of biological sample processing technology, specifically to a cytology specimen preprocessing system and method based on machine vision and intelligent decision-making. Background Technology

[0002] Pre-processing of cytological specimens is a crucial step in pathological diagnosis and molecular testing. The quality of this pre-processing directly determines the accuracy of subsequent slide preparation, diagnosis, and testing results. Cytological specimens (CS) are an important type of specimen for pathological diagnosis and molecular pathological testing, and CS pre-processing is a key step that determines the accuracy of pathological diagnosis and molecular testing. Currently, there are several clinical challenges in the pre-processing of cytology specimens. The volume of cytology specimens is increasing daily, and the pre-processing work in pathology cytology labs is heavy. There is an urgent need for high-quality automated specimen processing machines. Manual specimen processing carries the risk of errors, and there is an urgent need for automated identification systems to eliminate human-caused specimen errors. Manual operation is difficult to streamline, standardize, and regulate, which affects the quality of cytology slides and paraffin block preparation, leading to misdiagnosis. Currently, most cytology specimens do not have a fixed and disinfection procedure after collection, which may pose an infection risk to operators during the pre-processing process and may also cause the spread of airborne pathogens and contamination of groundwater resources. Although commercially available thin-layer liquid-based cytology instruments for preparing cytology specimens have improved the quality of slides for specimens such as sputum and pleural / peritoneal fluid, there is currently no automated pretreatment platform that can interface with thin-layer slide preparation equipment. If an automated integrated machine platform is directly integrated, there is a lack of design and equipment for automatically evaluating the specimen processing status. In actual scenarios, it still depends on the operator's experience. For samples with many impurities, filtering, multiple centrifugation, and multiple sputum / splitting red blood cells are used to achieve the optimal sample processing status to determine whether the next step of slide preparation can be carried out, which cannot realize the concept of automation. Traditional cytological pathological diagnosis only requires the preparation of cytological smears, smears, or liquid-based cytology slides. However, the preparation of cytological paraffin blocks has become an important part of the pretreatment of CS. Manual preparation of cytological paraffin blocks is time-consuming and labor-intensive, and the quality of the paraffin blocks needs to be improved. Summary of the Invention

[0003] This invention provides a cytology specimen pretreatment system and method based on machine vision and intelligent decision-making, which can effectively solve the problems mentioned in the background art, such as the need for manual processing of samples with many impurities, high requirements for the experience of operators, lack of fast and convenient evaluation methods, and the risk of sample contamination.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a cytology specimen pretreatment system based on machine vision and intelligent decision-making, comprising an operating platform, wherein the operating platform is provided with a first oscillation module, a centrifugation module, a visual recognition module, and a sample tube operation component; The first oscillation module is used to mix the sample liquid and reagents in the sample tube, the centrifugation module is used to centrifuge the mixed sample liquid, the visual recognition module is used to acquire images of the sample tube during the processing, and the sample tube operation component is used to perform sample tube transfer, opening and closing, and liquid suction and discharge operations. The first oscillation module and the sample tube operation component are respectively provided with two or more sets, each set can work independently, and when one set is in working state, the other set can be operated by the operator to pick up and put in the sample tube. The centrifugation module includes a turntable and a base. A rotary motor is fixedly connected to the base. The turntable is sleeved on the output shaft of the rotary motor. The turntable is provided with multiple insertion holes spaced around the output shaft. Each sample tube is reliably positioned in each insertion hole.

[0005] According to the above technical solution, the first oscillation module includes an oscillation plate and a base plate. The top surface of the oscillation plate has multiple fixed slide rails, and the oscillation plate is supported on the base plate by springs. An oscillation motor is installed on the base plate, and an eccentric wheel is sleeved on the output shaft of the oscillation motor. The sample tube is placed on the sample slot seat, and the bottom surface of the sample slot seat has a sliding groove that forms a snap-fit ​​limit with the fixed slide rails. The first oscillation module also includes a telescopic rod, which can apply force to each sample slot seat to make the sample slot seat slide out from the fixed slide rail.

[0006] According to the above technical solution, the sample tube operation component includes a gripper assembly and a pipetting assembly. The gripper assembly is equipped with a gripper lifting drive assembly to grip or screw the cap of the sample tube to realize the opening and closing of the cap. The pipetting assembly is equipped with a pipetting pump lifting drive assembly to drive the pipetting pump to move downward to insert the pipette tip for aspiration and dissipation of liquid in the sample tube. The operating platform is equipped with a pipette tip module and a pipette tip recycling bin. The pipette tip module contains multiple pipette tips for use by the pipetting assembly, and the pipette tip recycling bin is used to hold discarded pipette tips from the pipetting assembly.

[0007] According to the above technical solution, the operating platform further includes a motion driving device composed of a horizontal driving module and a vertical driving module, and the sample tube operating component is located on the horizontal driving module; The operating platform is also equipped with a light source, which is spaced apart from the visual recognition module, and the sample is placed between the light source and the visual recognition module.

[0008] According to the above technical solution, the operating platform is also provided with a sample retention module, the sample retention module has an empty sample tube inside, and the sample retention module is located adjacent to the first oscillation module; The operating platform is also equipped with a second oscillation module, which is located adjacent to the centrifugation module. The second oscillation module has a higher frequency and larger amplitude oscillation effect than the first oscillation module. The operating platform also includes a cell collection device, which can be placed at the bottom inside of the sample tube. After the collected cells are fixed with liquid, the centrifuged cells in the sample tube can be transferred to the target location. The cell collection device includes a filter bag and a lug on the top side of the filter bag, wherein the filter bag has a pore size of 0.05-1 mm.

[0009] According to the above technical solution, a method for preprocessing cytological specimens based on machine vision and intelligent decision-making includes the following steps: Step 1: Multimodal data acquisition and preprocessing; Step 2: Multimodal feature extraction and fusion model; Step 3: Calculation of dynamic quantization parameters; Step four: Reinforce learning-driven intelligent decision-making.

[0010] According to the above technical solution, in step one, multi-source data during the sample processing are collected comprehensively, and data quality is ensured through preprocessing; During image acquisition, a high-resolution multispectral camera system was used to continuously capture multi-angle images of the sample tube at a frequency of 5 frames per second during centrifugation. The camera covers the visible and near-infrared spectra. During image acquisition, the acquisition location is at the outlet side of the centrifugation module to ensure that the image covers the entire field of view of the sample tube, while the centrifugation time point is recorded. During the sensor data acquisition process, the physical parameters of the centrifugation module, as well as the frequency and amplitude of the first oscillation module, are collected in real time through sensing devices. The physical parameters include rotational speed, centrifugal force G value, and acceleration curve. At the same time, sample type metadata is recorded. The data is transmitted through IoT sensors with a sampling frequency of 100Hz. The specific preprocessing process includes image preprocessing and sensor data calibration.

[0011] According to the above technical solution, in step two, multi-source features are extracted and fused through a hybrid deep learning model; The model architecture design includes a visual TransformerViT branch and a CNN branch. The input to the visual TransformerViT branch is a stitched tensor of a multispectral image. The image is first divided into 16×16 pixel segments and input to the ViT encoder. ViT uses a self-attention mechanism to globally analyze cell layer morphology, mucus distribution, and impurity boundaries. The attention weights are visualized to identify key regions. The CNN branch uses a lightweight CNN in parallel to extract local texture features, focusing on capturing the microstructure of red blood cell clusters and the edge information of solid impurities. The CNN outputs a local feature map, which is fused with the global features of ViT through a cross-modal attention module.

[0012] According to the above technical solution, in step three, based on the fusion features, quantitative parameters are calculated and dynamic correction factors are introduced to make the parameters adapt to the processing process, including mucus residue index M, red blood cell residue R, cell layer morphology clarity C and centrifugation efficiency score E. The formula for calculating the mucus residue index M is as follows: M=(A m / A c )×K m ×F g ; A m and A c The pixel areas of the mucus and cell layers are respectively, and the semantic segmentation output is obtained through the ViT branch. K m This is a sample type correction coefficient; when the sensor detects a centrifugal force fluctuation greater than 10%, K m Automatically multiply by a correction factor, specifically 1.1-1.3; F g The centrifugal force correction factor is calculated using the following formula: F g =1 / (1+0.01×G), where G is the real-time centrifugal force; The formula for calculating erythrocyte residual R is as follows: R=(ΣA ri / A fov )×(1+α···T); ΣA ri The sum of the pixel areas of the red blood cell clusters is the target detection output through the CNN branch, where T is the centrifugation time and A is the red blood cell cluster area. fov α represents the total pixel area of ​​the image field of view, and α is the time decay factor. The formula for calculating the morphological clarity C of the cell layer is as follows: C=(G max -G min ) / Gavg ; G max G represents the maximum gray level of a pixel within a cell layer region. min G represents the minimum pixel grayscale value within the cell layer region. avg This represents the average pixel grayscale value within the cell layer region. The formula for calculating the centrifugation efficiency score E is as follows: E = β1·C + β2·(1-R) ​​- β3·M; The weights β1, β2, and β3 are initially set to 0.4, 0.3, and 0.3, respectively.

[0013] According to the above technical solution, step four involves replacing fixed rules with reinforcement learning to achieve dynamic decision generation. When setting up the decision engine, it includes a state space, an action space, and a reward function. The state space contains the real-time values ​​of parameters M, R, C, and E, as well as sensor data. The action space defines various operations, including adjusting centrifugation parameters, adding reagents, enabling filtration, and initiating secondary oscillation. The reward function is designed as follows: Reward = w1·E - w2·Time - w3·Cost, where Time is the processing time, Cost is the reagent consumption, and the weights w1, w2, and w3 are learned through training and the state is evaluated every 5 seconds during the decision-making process. Decision instructions are issued through the intelligent decision control system, driving the sample tube operation components to execute. After each decision is executed, a new round of data needs to be collected immediately to recalculate the parameters M, R, C, and E. If the parameters do not improve, an anomaly handling process is triggered. After manual intervention and data labeling, the data is used for model reinforcement learning to achieve a feedback loop.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By integrating the first oscillation module, centrifugation module, visual recognition module, sample tube operation components, and motion drive device, the system achieves automated mixing and centrifugation of sample solutions, thereby obtaining preliminary materials that meet the requirements of cytology specimens. This reduces the need for excessive manual intervention, lowers labor costs and the risk of infection for medical personnel during operation, and improves the efficiency of preliminary cytology specimen processing. It also requires less experience from operators and eliminates the risk of sample contamination. Furthermore, the system enables complete preparation of cytology specimens, including cytology paraffin blocks, thin-layer slides, and supernatant collection, increasing the opportunities for subsequent integration of cytology specimens with other testing platforms and enhancing the value of cytology specimens in pathological diagnosis.

[0015] 2. By acquiring images of the sample tubes during processing and calculating the mucus residue index M, red blood cell residue R, cell layer morphology clarity C, and centrifugation efficiency score E, the parameters are compared with preset thresholds to facilitate sample processing based on the results until the samples meet the qualification standards. This enables visual monitoring of each step in the slide preparation process. Samples that do not meet the standards undergo secondary or multiple optimization processes, thereby automating the pre-processing and effectively improving the qualification rate of samples before slide preparation. This significantly improves the standardization and efficiency of pre-processing quality, and reduces reliance on operator experience and the risk of biological contamination.

[0016] 3. By integrating visual images, centrifugation sensor data, and sample type metadata, the comprehensiveness of the evaluation is improved. A hybrid model of Visual TransformerViT and CNN is adopted to replace U-Net and YOLO. The global attention mechanism of ViT is used to capture subtle changes in cell layer morphology, enhance the identification of irregular impurities, and achieve precise control of the centrifugation process. Furthermore, by replacing fixed thresholds with reinforcement learning, the system can automatically optimize decision rules based on historical data to achieve dynamic parameters and reinforcement learning decisions, which significantly improves the standardization and intelligence of cytology pretreatment. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] In the attached diagram: Figure 1 This is a top view of the processing system of the present invention; Figure 2 This is a three-dimensional structural schematic diagram of the first oscillation module of the present invention; Figure 3 This is a three-dimensional structural schematic diagram of the centrifuge module of the present invention; Figure 4 This is a three-dimensional structural diagram of the processing system of the present invention; Figure 5 This is the present invention. Figure 4 A magnified view of a section at point A in the middle; Figure 6 This is a three-dimensional structural schematic diagram of the sample tube operation component of the present invention; Figure 7 This is a three-dimensional structural schematic diagram of the cell collection device of the present invention; Figure 8 This is a schematic diagram showing the cell collection device of the present invention placed inside a sample tube; Figure 9 This is a three-dimensional structural schematic diagram of the sample retention module of the present invention; Figure 10This is a flowchart of the processing method of the present invention; The diagram is labeled as follows: 1. Operating platform; 2. First oscillation module; 21. Oscillating plate; 211. Fixed slide rail; 22. Telescopic rod; 23. Spring; 24. Oscillating motor; 25. Base plate; 3. Centrifugation module; 31. Turntable; 32. Base; 41. Vision recognition module; 42. Light source; 5. Sample tube operation assembly; 51. Gripper assembly; 511. Gripper lifting drive assembly; 52. Pipette assembly; 521. Pipette pump lifting drive assembly; 61. Pipette tip module; 62. Pipette tip recovery chamber; 71. Lateral drive module; 72. Vertical drive module; 81. Sample retention module; 82. Second oscillation module; 9. Cell collection device; 91. Filter bag; 92. Lifting lug; 100. Sample tube; 101. Sample slot holder. Detailed Implementation

[0019] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0020] Example: Figure 1-9 As shown, the present invention provides a technical solution, a cytology specimen pretreatment system based on machine vision and intelligent decision-making, including an operation platform 1, on which a first oscillation module 2, a centrifugation module 3, a visual recognition module 41, and a sample tube operation component 5 are provided. The operation platform 1 also includes an outer cover covering it to form a relatively independent operation space. The outer cover has a corresponding sealing door that can be opened and closed, and a corresponding display screen is provided on the outer cover. The operator can view the captured photos through the display screen. The first oscillation module 2 is used to mix the sample solution and reagents in the sample tube 100 to separate the cells and liquid in the sample tube 100. The centrifugation module 3 is used to centrifuge the mixed sample solution. The visual recognition module 41 is used to acquire images of the sample tube 100 during the processing and to determine whether the sample after oscillation and centrifugation meets the requirements for cell suspension preparation before cytology slide preparation. The sample tube operation component 5 is used to perform the transfer, opening, closing and liquid aspiration and dissipation operations of the sample tube 100. The first oscillation module 2 and the sample tube operation component 5 are respectively set in two groups. Each group can work independently. When one group is in working state, the other group can be operated by the operator to pick up and put down the sample tube 100. The centrifuge module 3 includes a turntable 31 and a base 32. A rotary motor is fixedly connected to the base 32. The turntable 31 is sleeved on the output shaft of the rotary motor. The turntable 31 is provided with multiple insertion holes spaced around the output shaft. Each sample tube 100 is reliably positioned in each insertion hole.

[0021] The first oscillation module 2 includes an oscillation plate 21 and a base plate 25. The top surface of the oscillation plate 21 has multiple fixed slide rails 211, and the oscillation plate 21 is supported on the base plate 25 by springs 23. An oscillation motor 24 is installed on the base plate 25. An eccentric wheel is sleeved on the output shaft of the oscillation motor 24. The sample tube 100 is placed on the sample slot seat 101. The bottom surface of the sample slot seat 101 has a sliding groove that forms a snap-fit ​​limit with the fixed slide rails 211. The first oscillation module 2 also includes a telescopic rod 22, which can apply force to each sample slot seat 101 to make the sample slot seat 101 slide out from the fixed slide rail 211.

[0022] The sample tube operation assembly 5 includes a gripper assembly 51 and a pipetting assembly 52. ​​The gripper assembly 51 is equipped with a gripper lifting drive assembly 511 to screw the cap of the sample tube 100 to open and close the cap. The pipetting assembly 52 is equipped with a pipetting pump lifting drive assembly 521 to drive the pipetting pump to move downward to insert the pipette tip for aspiration and dissipation of liquid in the sample tube 100. The operating platform 1 is equipped with a pipette tip module 61 and a pipette tip recycling bin 62. The pipette tip module 61 contains multiple pipette tips for the pipetting assembly 52 to insert and use. The pipette tip recycling bin 62 is used to hold discarded pipette tips of the pipetting assembly 52.

[0023] The operating platform 1 also includes a motion driving device consisting of a horizontal driving module 71 and a vertical driving module 72, with the sample tube operating component 5 located on the horizontal driving module 71. The operating platform 1 is also equipped with a light source 42, which is spaced apart from the visual recognition module 41, and the sample is placed between the light source 42 and the visual recognition module 41.

[0024] The operating platform 1 is also equipped with a sample retention module 81, which contains an empty sample tube 100 and is located near the first oscillation module 2. The operating platform 1 is also equipped with a second oscillation module 82, which is located adjacent to the centrifugation module 3. The second oscillation module 82 has a higher frequency and larger amplitude oscillation effect than the first oscillation module 2, so as to quickly process samples with more mucus. The operating platform 1 also includes a cell collection device 9, which can be placed at the bottom inside of the sample tube 100. After the collected cells are fixed with liquid, the centrifuged cells in the sample tube 100 can be transferred to the target location. The cell collection device 9 includes a filter bag 91 and a lug 92 located on the top side of the filter bag 91. The filter bag 91 has a filter pore diameter of 0.5 mm.

[0025] like Figure 10As shown, a method for preprocessing cytological specimens based on machine vision and intelligent decision-making includes the following steps: Step 1: Multimodal data acquisition and preprocessing; Step 2: Multimodal feature extraction and fusion model; Step 3: Calculation of dynamic quantization parameters; Step four: Reinforce learning-driven intelligent decision-making.

[0026] Based on the above technical solution, the first step is to comprehensively collect multi-source data during the sample processing and ensure data quality through preprocessing, laying the foundation for subsequent analysis. During image acquisition, a high-resolution multispectral camera system was used to continuously capture multi-angle images of the sample tube at a frequency of 5 frames per second during centrifugation. The camera covered the visible and near-infrared spectra to enhance the contrast between mucus, red blood cells and cell layers. Near-infrared light can penetrate the mucus layer and clearly show the distribution of cells at the bottom. During image acquisition, the acquisition position is located at the outlet side of centrifugation module 3 to ensure that the image covers the entire field of view of sample tube 100 and avoids occlusion. At the same time, the centrifugation time point is recorded to facilitate alignment with sensor data. During the sensor data acquisition process, the physical parameters of the centrifugation module 3, as well as the frequency and amplitude of the first oscillation module 2, are collected in real time through the sensing device. The physical parameters include rotational speed, centrifugal force G value, and acceleration curve. At the same time, sample type metadata, including the preset viscosity coefficient Km, is entered. Data is transmitted via an IoT sensor with a sampling frequency of 100Hz to ensure synchronization with the image frame timestamp; the IoT sensor is an encoder. The specific preprocessing process includes image preprocessing and sensor data calibration; Image preprocessing involves adaptive histogram equalization and nonlocal mean denoising of the original image to reduce uneven illumination and noise interference. Then, multispectral registration is performed to ensure pixel-level alignment of images in different bands. Sensor data calibration involves using low-pass filtering to smooth the centrifugal force data, eliminating mechanical vibration noise, and normalizing all data to the [0,1] interval to facilitate multimodal fusion. The process generates a time-aligned multimodal dataset, including registered image sequences, calibrated sensor data, and metadata.

[0027] Based on the above technical solution, in step two, multi-source features are extracted and fused through a hybrid deep learning model to improve the comprehensiveness and robustness of feature representation. The model architecture design includes a visual TransformerViT branch and a CNN branch; The input to the Vision TransformerViT branch is a stitched tensor of a multispectral image with a size of 1024×1024×6, corresponding to 3 visible light and 3 near-infrared channels. The image is first divided into 16×16 pixel segments and then input into the ViT encoder. ViT utilizes a self-attention mechanism to globally analyze cell layer morphology, mucus distribution, and impurity boundaries, with particular optimization of long-term dependence on irregular regions. Attention weight visualization can identify key regions. The CNN branch uses a lightweight CNN in parallel to extract local texture features, focusing on capturing the microstructure of red blood cell clusters and edge information of solid impurities. The CNN outputs local feature maps, which are fused with the global features of ViT through a cross-modal attention module. The feature fusion module concatenates the global feature vector of ViT, the local feature map of CNN, and sensor data into a multimodal feature vector. During fusion, a weighted summation is used, and the weights are dynamically calculated by the attention mechanism to ensure that high-contribution features dominate the decision-making process. In addition, adversarial training is used to enhance the model's generalization ability. Rare samples are simulated by using a generative adversarial network (GAN). The loss function combines Dice loss and focus loss. The model needs to be fine-tuned with newly collected data every 24 hours to ensure continuous adaptation to sample variation.

[0028] Based on the above technical solution, in step three, based on the fusion features, quantitative parameters are calculated and dynamic correction factors are introduced to make the parameters adaptively change with the processing process, including mucus residue index M, red blood cell residue R, cell layer morphology clarity C and centrifugation efficiency score E. The formula for calculating the mucus residue index M is as follows: M=(A m / A c )×K m ×F g ; A m and A c The pixel areas of the mucus and cell layers are respectively, and the semantic segmentation output is obtained through the ViT branch. K m This is a sample type correction coefficient; when the sensor detects a centrifugal force fluctuation greater than 10%, K m Automatically multiply by a correction factor, which is 1.2. F g The centrifugal force correction factor is calculated using the following formula: F g =1 / (1+0.01×G), where G is the real-time centrifugal force. A high G value can compress the mucus, thereby reducing the estimated M value. The M value is updated every 2 seconds to dynamically reflect the mucus residue. The formula for calculating erythrocyte residual R is as follows: R=(ΣA ri / A fov )×(1+α···T); ΣA ri The sum of the pixel areas of the red blood cell clusters is the target detection output through the CNN branch, where T is the centrifugation time and A is the red blood cell cluster area. fov The total pixel area of ​​the image field of view is α, which is the time decay factor. The initial value is 0.05. It is optimized by learning from historical data. During long-term centrifugation, α automatically increases to avoid overprocessing. The formula for calculating the morphological clarity C of the cell layer is as follows: C=(G max -G min ) / G avg ; G max G represents the maximum gray level of a pixel within a cell layer region. min G represents the minimum pixel grayscale value within the cell layer region. avg The average gray value of pixels within the cell layer region is calculated by adding gradient-assisted calculation, using ViT to extract optical flow features of the cell layer boundary, calculating gradient magnitude, and fusing it into the C value. The C value combines gray-level statistics and boundary sharpness. The formula for calculating the centrifugation efficiency score E is as follows: E = β1·C + β2·(1-R) ​​- β3·M; The weights β1, β2, and β3 are initially set to 0.4, 0.3, and 0.3, respectively. The E value ranges from [0,1]. A value ≥0.7 indicates that the result is acceptable, while a value <0.6 triggers a decision adjustment.

[0029] Based on the above technical solution, in step four, reinforcement learning is used to replace fixed rules to achieve dynamic decision generation and optimize processing efficiency. When setting up the decision engine, the state space, action space and reward function are included. The state space contains the real-time values ​​of parameters M, R, C, and E, as well as sensor data. The action space defines various operations, including adjusting centrifugation parameters, adding reagents, enabling filtration, and initiating secondary oscillation. The reward function is designed as Reward = w1·E - w2·Time - w3·Cost; Where Time is the processing time, Cost is the reagent consumption, and weights w1, w2, and w3 are learned through training. During the decision-making process, the state is evaluated every 5 seconds. If E < 0.6, the reinforcement learning agent selects the action. The decision-making instructions are issued through the intelligent decision control system, driving the sample tube operation component 5 to execute. After each decision is executed, a new round of data needs to be collected immediately, including image data and sensor data, and the parameters M, R, C, and E are recalculated. If the parameters do not improve, the abnormal handling process is triggered, that is, the current sample is paused, and the data is manually labeled and used for model reinforcement learning to achieve feedback loop.

[0030] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cytology specimen preprocessing system based on machine vision and intelligent decision-making, comprising an operating platform (1), characterized in that: The operating platform (1) is equipped with a first oscillation module (2), a centrifugation module (3), a visual recognition module (41), and a sample tube operation component (5). The first oscillation module (2) is used to mix the sample liquid and reagent in the sample tube (100), the centrifugation module (3) is used to centrifuge the mixed sample liquid, the visual recognition module (41) is used to acquire images of the sample tube (100) during the processing, and the sample tube operation component (5) is used to perform the transfer, opening, closing and liquid suction and discharge operations of the sample tube (100). The first oscillation module (2) and the sample tube operation component (5) are respectively provided with two or more sets, each set can work independently, and when one set is in working state, the other set can be operated by the operator to pick up and put down the sample tube (100); The centrifugation module (3) includes a turntable (31) and a base (32). A rotary motor is fixedly connected to the base (32). The turntable (31) is sleeved on the output shaft of the rotary motor. The turntable (31) is provided with multiple insertion holes spaced around the output shaft. Each sample tube (100) is reliably positioned in each insertion hole.

2. The cytology specimen pretreatment system based on machine vision and intelligent decision-making according to claim 1, characterized in that: The first oscillation module (2) includes an oscillation plate (21) and a base plate (25). The top surface of the oscillation plate (21) has multiple fixed slide rails (211), and the oscillation plate (21) is supported on the base plate (25) by springs (23). An oscillation motor (24) is installed on the base plate (25). An eccentric wheel is sleeved on the output shaft of the oscillation motor (24). The sample tube (100) is placed on the sample slot seat (101). The bottom surface of the sample slot seat (101) has a sliding groove that forms a snap-fit ​​limit with the fixed slide rails (211). The first oscillation module (2) also includes a telescopic rod (22), which can apply force to each sample slot seat (101) to make the sample slot seat (101) slide out from the fixed slide rail (211).

3. The cytology specimen pretreatment system based on machine vision and intelligent decision-making according to claim 1, characterized in that: The sample tube operation assembly (5) includes a gripper assembly (51) and a pipetting assembly (52). The gripper assembly (51) is equipped with a gripper lifting drive assembly (511) to grip or screw the cap of the sample tube (100) to open and close the cap. The pipetting assembly (52) is equipped with a pipetting pump lifting drive assembly (521) to drive the pipetting pump to move downward to insert the pipette tip for aspiration and dissipation of liquid in the sample tube (100). The operating platform (1) is provided with a pipette tip module (61) and a pipette tip recycling bin (62). The pipette tip module (61) has multiple pipette tips for the pipetting assembly (52) to insert and use. The pipette tip recycling bin (62) is used to hold discarded pipette tips of the pipetting assembly (52).

4. The cytology specimen pretreatment system based on machine vision and intelligent decision-making according to claim 1, characterized in that: The operating platform (1) also includes a motion driving device consisting of a horizontal driving module (71) and a vertical driving module (72), and the sample tube operating component (5) is located on the horizontal driving module (71); The operating platform (1) is also equipped with a light source (42), which is spaced apart from the visual recognition module (41), and the sample is placed between the light source (42) and the visual recognition module (41).

5. The cytology specimen pretreatment system based on machine vision and intelligent decision-making according to claim 1, characterized in that: The operating platform (1) is also provided with a sample retention module (81), which has an empty sample tube (100) inside. The sample retention module (81) is located adjacent to the first oscillation module (2). The operating platform (1) is also provided with a second oscillation module (82), which is located adjacent to the centrifugation module (3). The second oscillation module (82) has a higher frequency and greater amplitude oscillation effect than the first oscillation module (2). The operating platform (1) also includes a cell collection device (9), which can be placed at the bottom inside of the sample tube (100). After the collected cells are fixed with liquid, the cells centrifuged from the sample tube (100) can be transferred to the target location. The cell collection device (9) includes a filter bag (91) and a lug (92) on the top side of the filter bag (91), wherein the filter bag (91) has a filter pore diameter of 0.05-1 mm.

6. A method for preprocessing cytological specimens based on machine vision and intelligent decision-making, used to process cytological specimens preprocessing systems based on machine vision and intelligent decision-making as described in any one of claims 1-5, characterized in that: Includes the following steps: Step 1: Multimodal data acquisition and preprocessing; Step 2: Multimodal feature extraction and fusion model; Step 3: Calculation of dynamic quantization parameters; Step four: Reinforce learning-driven intelligent decision-making.

7. The method for preprocessing cytological specimens based on machine vision and intelligent decision-making according to claim 6, characterized in that: Step one involves comprehensively collecting multi-source data during the sample processing and ensuring data quality through preprocessing. During image acquisition, a high-resolution multispectral camera system was used to continuously capture multi-angle images of the sample tube (100) at a frequency of 5 frames per second during centrifugation. The camera covers the visible and near-infrared spectra. During image acquisition, the acquisition location is located at the outlet side of the centrifugation module (3) to ensure that the image covers the entire field of view of the sample tube (100) and the centrifugation time point is recorded at the same time; During the process of acquiring sensor data, the physical parameters of the centrifugation module (3) and the frequency and amplitude of the first oscillation module (2) are collected in real time through the sensing device. The physical parameters include rotation speed, centrifugal force G value, and acceleration curve. At the same time, sample type metadata is recorded. The data is transmitted through the Internet of Things sensor and the sampling frequency is 100Hz. The specific preprocessing process includes image preprocessing and sensor data calibration.

8. The method for preprocessing cytological specimens based on machine vision and intelligent decision-making according to claim 6, characterized in that: In step two, multi-source features are extracted and fused using a hybrid deep learning model. The model architecture design includes a Visual TransformerViT branch and a CNN branch. The input to the Visual TransformerViT branch is a stitched tensor of a multispectral image. The image is first divided into 16×16 pixel segments, which are then input to the ViT encoder. ViT uses a self-attention mechanism to globally analyze cell layer morphology, mucus distribution, and impurity boundaries. The attention weights are visualized to identify key regions. The CNN branch uses a lightweight CNN in parallel to extract local texture features, focusing on capturing the microstructure of red blood cell clusters and the edge information of solid impurities. The CNN outputs a local feature map, which is fused with the global features of ViT through a cross-modal attention module.

9. A method for preprocessing cytological specimens based on machine vision and intelligent decision-making according to claim 6, characterized in that: In step three, based on the fusion features, quantitative parameters are calculated and dynamic correction factors are introduced to make the parameters adaptively change with the processing process, including mucus residue index M, red blood cell residue R, cell layer morphology clarity C, and centrifugation efficiency score E. The formula for calculating the mucus residue index M is as follows: M=(A m / A c )×K m ×F g ; A m and A c The pixel areas of the mucus and cell layers are respectively, and the semantic segmentation output is obtained through the ViT branch. K m This is a sample type correction coefficient; when the sensor detects a centrifugal force fluctuation greater than 10%, K m Automatically multiply by a correction factor, specifically 1.1-1.3; F g The centrifugal force correction factor is calculated using the following formula: F g =1 / (1+0.01×G), where G is the real-time centrifugal force; The formula for calculating erythrocyte residual R is as follows: R=(ΣA ri / A fov )×(1+α···T); ΣA ri The sum of the pixel areas of the red blood cell clusters is the target detection output through the CNN branch, where T is the centrifugation time and A is the red blood cell cluster area. fov α represents the total pixel area of ​​the image field of view, and α is the time decay factor. The formula for calculating the morphological clarity C of the cell layer is as follows: C=(G max -G min ) / G avg ; G max G represents the maximum gray level of a pixel within a cell layer region. min G represents the minimum pixel grayscale value within the cell layer region. avg This represents the average pixel grayscale value within the cell layer region. The formula for calculating the centrifugation efficiency score E is as follows: E = β1·C + β2·(1-R) ​​- β3·M; The weights β1, β2, and β3 are initially set to 0.4, 0.3, and 0.3, respectively.

10. A method for preprocessing cytological specimens based on machine vision and intelligent decision-making according to claim 6, characterized in that: Step four involves replacing fixed rules with reinforcement learning to achieve dynamic decision generation. When setting up the decision engine, it includes a state space, an action space, and a reward function. The state space contains the real-time values ​​of parameters M, R, C, and E, as well as sensor data. The action space defines various operations, including adjusting centrifugation parameters, adding reagents, enabling filtration, and initiating secondary oscillation. The reward function is designed as follows: Reward = w1·E - w2·Time - w3·Cost, where Time is the processing time, Cost is the reagent consumption, and the weights w1, w2, and w3 are learned through training and the state is evaluated every 5 seconds during the decision-making process. The decision instruction is issued through the intelligent decision control system, which drives the sample tube operation component (5) to execute. After each decision is executed, a new round of data needs to be collected immediately to recalculate the parameters M, R, C, and E. If the parameters are not improved, the abnormal handling process is triggered. After manual intervention and data labeling, the data is used for model reinforcement learning to achieve feedback loop.

Images