UBM image quantitative analysis method and system for silicone oil emulsification evaluation

Abstract

The invention discloses a UBM image quantitative analysis method and system for silicone oil emulsification evaluation, belongs to the field of medical image processing, and aims to objectively quantify emulsification-related acoustic signs in a silicone oil filling postoperative UBM image. The method comprises the following steps: communicating with UBM equipment, monitoring gain and probe perpendicularity, and outputting a control / prompt instruction for standardized acquisition; desensitizing the image and masking a non-scanning area; extracting the characteristics of floating particle density, artifact intensity, high-echo particle aggregation region area proportion, tissue infiltration signal-to-noise ratio, corneal endothelial post-deposition density and the like; and fusing the center and each orientation feature, introducing a space weight matrix to calculate a weighted acoustic distribution feature value ADCV, and outputting the ADCV and a structured quantization result for clinical evaluation and follow-up comparison.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and computer-aided diagnostic technology, and in particular to a UBM image quantitative analysis method and system for assessing silicone oil emulsification. Background Technology

[0002] Ultrasonic biomicroscopy (UBM) is a crucial imaging tool for observing the fine structures of the anterior segment and has significant application value in the postoperative evaluation of silicone oil tamponade. Silicone oil, a commonly used long-lasting tampon in vitreoretinal surgery, can emulsify and form micron-sized droplets if left in the eye for extended periods, potentially inducing complications such as secondary glaucoma and corneal endothelial decompensation. Therefore, repeatable and comparable assessments of the degree of silicone oil emulsification are necessary to assist clinicians in determining the timing of intervention and follow-up strategies.

[0003] Existing UBM assessment methods for silicone oil emulsification have generally followed two paths: The first is subjective qualitative assessment, which relies primarily on physicians' visual observation of UBM images and judgments based on experienced descriptions such as "punctate echoes" and "pine needle-like artifacts." This approach is inherently subjective and difficult to compare across operators and time periods. The second is a semi-quantitative scoring system. To reduce subjectivity, this system proposes a 0 / 1 discrete scoring system based on multiple emulsification signs, which is then summed to obtain a total score (e.g., the total score is composed of several sign items from a horizontal wide-view image and infiltration sign items from eight narrow-view images). This system often employs double-blind interpretation and disagreement arbitration procedures to obtain the final score.

[0004] Although the above semi-quantitative system is an improvement over purely subjective evaluation, the following technical bottlenecks still exist: (1) Limited quantification accuracy: Existing scores are mostly "present / absent" or graded discrete scores, which are difficult to reflect the subtle changes in continuous physical quantities such as particle density, artifact intensity, and the proportion of high echo particle aggregation area, thus lacking sensitivity to disease progression or efficacy monitoring.

[0005] (2) Lack of automation constraints in the acquisition process: The lack of real-time monitoring and closed-loop control of key acquisition conditions such as UBM device gain and probe angle / verticality results in image quality being greatly affected by the operator and poor comparability of data from different sources.

[0006] (3) Insufficient clinical relevance: Existing scoring models usually treat different orientations equally, without explicitly considering the spatial distribution differences of emulsified silicone oil under the influence of gravity, resulting in insufficient correlation between the results and the real clinical risk.

[0007] (4) Lack of compliance preprocessing: Existing processes usually do not integrate privacy protection preprocessing steps such as desensitization of medical image data, making it difficult to meet the compliance requirements of actual data flow and application.

[0008] (5) Lack of an integrated automatic quantification scheme for UBM silicone oil emulsion-specific features: General image processing methods have not formed a closed-loop process from ensuring the consistency of quality at the source of acquisition to multi-dimensional feature extraction and comprehensive quantification for features such as gravity-dependent distribution, characteristic vertical artifacts, and small hyperechoic spots.

[0009] Therefore, there is an urgent need for an assessment method and system that can objectively, continuously, and comparatively quantify the acoustic characteristics related to silicone oil emulsification while ensuring the consistency of data acquisition quality, and can reflect the differences in directional risks and meet data compliance requirements. Summary of the Invention

[0010] The technical problem to be solved by this invention is: in the UBM examination after silicone oil filling, how to transform the acoustic signs related to emulsification from manual discrete scoring into repeatable continuous physical quantitative results while ensuring the consistency of acquisition conditions and compliance with data desensitization, and how to make the quantitative results reflect the differences in clinical risk in different anatomical orientations (especially the inferior orientation).

[0011] The technical solution adopted by this invention to solve its technical problem is: A UBM image quantitative analysis method for silicone oil emulsification assessment includes the following steps: S0 Image Acquisition Standardization Closed-Loop Control: Communicates with the Ultrasonic Biological Microscope (UBM) device to acquire acquisition parameters and / or image quality indicators in real time; when the acquisition parameters and / or quality indicators do not meet the preset standards, control commands and / or interactive prompts are generated to constrain the acquisition process, thereby obtaining a target image that meets the standardization requirements. Preferably, the preset standard includes a gain value G in the range of 79dB to 81dB and / or a corneal echo width W less than or equal to a preset threshold, and can determine the output lock / unlock and allow save commands based on consecutive frames.

[0012] S1 Image Preprocessing: The target image is subjected to data desensitization processing, and the non-effective scan area is masked to obtain the effective scan area image; S2 Multidimensional Acoustic Feature Extraction: Perform morphological operations, spot detection, threshold segmentation, texture analysis, and / or signal-to-noise ratio calculation on the effective scan area image to extract continuous physical quantization features for quantifying the distribution of the medium; the continuous physical quantization features include at least one of the following: anterior chamber floating particle density, intensity of retro-endothelial multiple reflection artifacts, area ratio of hyperechoic particle aggregation region, tissue infiltration signal-to-noise ratio, and retro-endothelial deposition density; Preferably, the density of anterior chamber airborne particles is obtained through anterior chamber ROI construction, top cap transformation enhancement, multi-scale LoG response extreme value screening, grayscale and roundness threshold screening, and density calculation based on pixel spacing to convert the true area of ​​the ROI; artifact intensity is obtained through artifact analysis area localization, Sobel gradient enhancement, and grayscale co-occurrence matrix (GLCM) contrast calculation; tissue infiltration is calculated through target / background ROI statistics and determined according to SNR=(μ_target−μ_bg) / σ_bg; the area ratio of hyperechoic particle aggregation region is obtained through threshold segmentation, morphological optimization, and convex hull area statistics and calculated according to the area ratio; deposition density is obtained through narrowband ROI construction along the corneal endothelial boundary, local adaptive threshold segmentation, morphological denoising, and density calculation.

[0013] S3 Spatial Weighted Quantization Evaluation: The continuous physical quantization features obtained in step S2 are normalized and fused to obtain the central region feature value P_center and the feature values ​​P_i in each direction; a spatial weight matrix is ​​constructed to assign weight coefficients w_i to different directions, and the weighted acoustic distribution feature value ADCV is calculated; Preferably, the target image includes one wide-view image and eight narrow-view images from different orientations, wherein the orientations include at least 1:30, 3:00, 4:30, 6:00, 7:30, 9:00, 10:30, and 12:00; the spatial weight matrix assigns a higher weight coefficient to the lower orientation set S_inf than to other orientations, wherein S_inf includes at least the 6:00 orientation, and optionally further includes the adjacent 4:30 and / or 7:30 orientations; the ADCV is calculated according to the following formula: ADCV = 100×( α·P_center + β·( Σ(w_i·P_i) / Σw_i ) ), where α+β=1.

[0014] S4 Result Output: Output the ADCV and / or the structured quantization results associated with it; preferably, the structured quantization results include ADCV, each continuous physical quantization feature value, acquired quality control information, and risk classification prompts based on ADCV.

[0015] This invention also provides a UBM image quantization analysis system for silicone oil emulsification evaluation, comprising: an acquisition control and monitoring module, an image preprocessing module, a feature extraction module, a spatial weighted quantization module, and an output module; wherein, the acquisition control and monitoring module is used to perform image acquisition standardization closed-loop control in step S0, the image preprocessing module is used to perform image desensitization and effective scanning area masking processing in step S1, the feature extraction module is used to perform multi-dimensional acoustic feature extraction in step S2, the spatial weighted quantization module is used to perform spatial weighted quantization evaluation and calculate ADCV in step S3, and the output module is used to output the results in step S4.

[0016] The beneficial effects of this invention are: Compared with the prior art, the present invention has at least the following effects: (1) By standardizing closed-loop control of image acquisition, the gain parameters and probe verticality related quality indicators during the UBM acquisition process are monitored in real time, and control commands and / or interactive prompts are output when the preset standards are not met, so that the input images meet the consistency requirements, providing a stable data foundation for subsequent quantitative analysis and improving the comparability and repeatability of acquisition results from different operators and at different time points.

[0017] (2) Data desensitization and effective scanning area masking preprocessing are integrated before the image enters the analysis process. This can remove patient identification information and interference from non-scanning areas while retaining necessary metadata such as pixel spacing, providing compliant and reusable data input conditions for quantitative calculation.

[0018] (3) For the acoustic signs related to silicone oil emulsification, a multi-dimensional continuous physical quantitative feature extraction mechanism was constructed, which can output quantitative results such as the density of airborne particles, the intensity of artifacts, the area ratio of high echo particle aggregation region, the signal-to-noise ratio of tissue infiltration and the density of corneal endothelial deposits. This transforms the assessment from discrete manual scoring into a calculable and traceable continuous indicator, which is convenient for follow-up comparison and trend analysis.

[0019] (4) By normalizing and integrating the feature values ​​of the central region with the feature values ​​of each direction, and introducing a spatial weight matrix to give higher weight to the high-risk directions below, the risk differences of different anatomical directions can be reflected in the comprehensive index, so that the quantitative results match the gravity-dependent distribution characteristics observed in clinical practice.

[0020] (5) By calculating the weighted acoustic distribution characteristic value ADCV and outputting the structured quantitative results, a result report including ADCV, characteristic physical quantities of each dimension, collection quality control information and risk classification prompts can be generated, providing a data carrier that can be directly used for clinical assessment and follow-up management. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall process of the medium acoustic distribution quantitative analysis method of the present invention. Figure 2 This is a schematic diagram of the orientation division of the spatial weight matrix of the present invention and the lower high-weight orientation, wherein the lower orientation set S_inf includes at least the 6:00 orientation, and preferably further includes the adjacent 4:30 and / or 7:30 orientations; Figure 3 A standardized closed-loop control flowchart for image acquisition; Figure 4 A system module structure diagram for implementing the method of the present invention; Figure 5 This is a schematic diagram of typical UBM images and signs of silicone oil emulsification, where A represents anterior chamber floating particles, B represents corneal endothelial deposits, C represents ghosting images / pine needle artifacts, and D represents anterior chamber hypopyon-like aggregation; white arrows, black arrows, and *, △, and ☆ are used to mark different signs of tissue infiltration or aggregation; AC represents the anterior chamber, SO represents silicone oil, and C represents the cornea. Detailed Implementation

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings, so that those skilled in the art can implement it accordingly. This embodiment is used to explain the method and system described in claims 1 to 10, and is not intended to limit the scope of protection.

[0023] Example 1: A UBM image quantification analysis method for silicone oil emulsification assessment like Figure 1 As shown, the method in this embodiment includes steps S0 to S4.

[0024] S0 image acquisition standardized closed-loop control (see...) Figure 3 ) To reduce the impact of operator differences on subsequent quantization, this embodiment implements closed-loop quality control for the UBM acquisition process. The system reads the current gain value G (in dB) in real time via the DICOM protocol or the device SDK, setting the standard range to [79, 81] dB. When G < 79 dB or G > 81 dB is detected, the system sends a LOCK command to the device to pause image freezing and saving. If the UBM device does not support lock / unlock commands, the acquisition software will prevent the current frame from being written to the storage medium or mark it as an invalid frame, and prompt the operator to re-acquire. Once the acquisition parameters and image quality indicators meet the preset standards, freezing and saving are allowed, and the subsequent analysis process begins. The interface prompts the operator to adjust the gain to near the standard range. Meanwhile, the system performs verticality feedback frame by frame on the acquired video stream: Canny edge detection is applied to the central region of the image to extract corneal echo lines, grayscale profiles are obtained along their normal direction and the full width at half maximum (FWHM) is fitted and calculated to obtain the corneal echo line width W (unit pixel); W_max is set to 3 pixels, when the average value of W for 5 consecutive frames exceeds W_max, the probe is determined to be skewed and an adjustment is prompted, and when W is continuously lower than W_max and tends to the minimum value, saving is allowed.

[0025] The standardized acquisition sequence is as follows: first, acquire one horizontal wide-view image, followed by eight narrow-view images in sequence, with the orientation sequence being 1:30, 3:00, 4:30, 6:00, 7:30, 9:00, 10:30, and 12:00; each orientation image must meet the above preset standards for G and W before it can be frozen and saved.

[0026] S1 Image Preprocessing S1a Data Anonymization: The system retrieves the 9 DICOM files (1 wide view + 8 narrow view) saved in S0, parses their metadata, and deletes / erases data elements containing Personally Identifiable Information (PII), including (0010,0010) patient name, (0010,0020) patient ID, (0010,0030) date of birth, (0008,0020) examination date, (0008,0080) institution name, (0008,0090) referring physician, etc.; the above fields are replaced with a uniform anonymous identifier (e.g., "SiliconeOil_001"), while retaining non-PII metadata necessary for analysis, such as row and column count (0028,0010) / (0028,0011), pixel pitch (0028,0030), grayscale interpretation (0028,0103), etc.; the anonymized dataset is stored in an access-controlled directory.

[0027] S1b Effective Region Mask: The DICOM image is converted into a grayscale matrix I, a grayscale histogram is calculated, and automatic thresholding using Otsu's method is used to obtain a binary image. The largest connected region in the binary image is selected as the effective sector scan area, generating a binary mask M (effective area is 1, others are 0). Non-scanning dark areas and non-intraocular data such as image edge scales, text, and logos are forced to zero through pixel-by-pixel multiplication I_masked = I*M, and the masked image I_masked is output for use by S2. When the largest connected region in the binary image does not satisfy a preset shape prior (e.g., area ratio, sector boundary curvature, or coverage area located at the image center), the largest connected region is determined not to be an effective sector scan area. In this case, a preset sector template mask can be used instead, or candidate connected regions can be reordered and selected according to shape priors to obtain an effective scan area mask. Optionally, image edge scales / text areas can be removed or morphologically denoised before mask generation to improve the stability of the largest connected region determination.

[0028] S2 Multidimensional Acoustic Feature Extraction This embodiment performs continuous physical quantification or determination of the following five core features: anterior chamber floating particle density, intensity of retro-endothelial multiple reflection artifacts, area ratio of hyperechoic particle aggregation region, tissue infiltration signal-to-noise ratio determination, and retro-endothelial deposition density.

[0029] Before describing the automation algorithm, it is necessary to first clarify the specific object to be quantified and analyzed in this invention. For example... Figure 5 The image shown is a typical UBM image of a silicone oil-emulsified eye, clearly marking various pathological acoustic features: Area A (arrow): This shows the fore-house suspended particles that need to be automatically counted and their density calculated by step S2.1 of this invention.

[0030] Region B: This section illustrates step S2.5 of the present invention, which aims to detect and quantify post-endothelial deposits of the cornea.

[0031] Region C: The image shown is a ghost image (vertical artifact) formed by multiple reflections at the silicone oil-aqueous interface, the intensity of which will be quantitatively evaluated by step S2.2 of this invention.

[0032] Region D: The area of ​​high-echoic particle aggregation indicated by the white arrow, i.e., hypopyon-like aggregation, the area ratio of which will be calculated by step S2.3; The areas marked by the black arrow and the symbols *, △, and ☆ correspond to the infiltration of the anterior chamber angle, the anterior surface of the iris, the posterior surface of the iris, and the ciliary body, respectively. The abnormal echo enhancement in these areas will be automatically determined by step S2.4 by calculating the local signal-to-noise ratio (SNR).

[0033] Figure 5 This intuitively defines the objective of all subsequent algorithmic processing in this invention. The following will elaborate on how computer vision algorithms can be used to... Figure 5 The various symptoms shown are subjected to automated and quantitative feature extraction.

[0034] S2.1 Anterior chamber suspended particle density ROI Construction: Input a single UBM image (wide or narrow view), perform edge detection (e.g., Canny) to obtain a set of candidate edge points; extract edge points corresponding to the posterior corneal echo band in the upper region of the image, and use RANSAC to perform quadratic curve fitting on the edge points to obtain the corneal endothelial boundary curve L_upper. In the anterior chamber region below L_upper, prioritize extracting and fitting the anterior iris boundary curve L_iris; when the upper edge of the anterior capsule of the lens can be reliably identified in the image, the "reliable identification" can be obtained based on a comprehensive judgment of the continuous length of the boundary, the edge gradient intensity, and the curve fitting residual. For example, when the continuous length of the candidate boundary exceeds a preset proportion and the fitting residual is less than a preset threshold, the upper edge boundary curve of the anterior capsule of the lens is determined to be usable. Further extract and fit the upper edge curve of the anterior capsule of the lens L_lens. The one above L_iris and L_lens is taken as the lower boundary curve of the anterior chamber L_lower; connect the endpoints of L_upper and L_lower to form a closed polygonal region, which is the region of interest (ROI) of the anterior chamber. If the extraction of the L_iris boundary curve of the anterior iris fails or the fitting is unstable, the strip ROI located within a preset depth range below L_upper is used as the anterior chamber ROI, or the lower boundary of the anterior chamber is determined by manual correction to ensure that subsequent particle counting and density calculation can be implemented.

[0035] Top Hat Transformation: Define a circular structuring element B with a radius of 5 pixels. Perform a morphological opening operation Open(I,B) on the original image I within the ROI and calculate I_tophat = I − Open(I,B) to suppress slowly changing backgrounds and enhance grainy highlights.

[0036] Multi-scale LoG spot detection: Considering that the diameter of emulsified silicone oil particles is mainly distributed in the range of 10 to 50 μm, and the image scale of this system is, for example, 20 μm / pixel, corresponding to a spot radius of about 0.5 to 2.5 pixels; using the matching relationship of σ≈r / √2, LoG filtering at three scales of σ1=1.5, σ2=2.5, and σ3=3.5 is used to process I_tophat in parallel, and the maximum value of the response at the three scales is taken as the final spot intensity map.

[0037] Spot filtering: Before threshold judgment, pixel values ​​can be linearly mapped to the 8-bit grayscale range [0,255] (or normalized to [0,1]) based on DICOM grayscale interpretation and bit depth to ensure consistency of threshold conditions under different devices / different bit depths. A candidate connected component set is generated on the spot intensity map; intensity filtering is performed (spot intensity > 120, determined based on the grayscale distribution of 100 positive regions, which can filter more than 95% of background noise), and shape filtering is performed (circularity = 4π*Area / Perimeter^2, retaining roundness > 0.8 to exclude linear artifacts).

[0038] Density calculation: Count the number of connected components N after filtering; convert the actual area A_mm² of ROI based on the DICOM pixel spacing label (0028,0030); calculate Density=N / A_mm², unit: units / mm².

[0039] S2.2 Intensity of Retroendothelial Multiple Reflection Artifacts Step 1: Artifact Analysis Area Location: Using the corneal endothelial boundary L_upper as a reference, offset 1mm (converted to pixel displacement according to the pixel scale, for example, about 50 pixels) inwards to obtain a band-shaped area parallel to L_upper as the artifact analysis area.

[0040] Step 2 Texture Enhancement: Perform a Sobel Y convolution on the grayscale image within the artifact analysis area, with a kernel of Gy=[-1 -2 -1; 0 0 0; +1 +2 +1], to specifically enhance the vertical stripe texture.

[0041] Step 3 Texture Quantization: Construct a gray-level co-occurrence matrix (GLCM) on the Sobel Y result, set the pixel pair distance d=1 and the orientation angle θ=0°, and calculate the contrast Contrast=Σ(ij)^2*P(i,j), where P(i,j) is the probability that gray levels i and j co-occur under a given orientation and distance.

[0042] Step 4 Intensity calibration: Output Contrast as the artifact intensity quantification index Artifact_Strength (a continuous scalar, the larger the value, the more significant the artifact).

[0043] S2.3 Percentage of area of ​​high-echo particle aggregation region Step 1 Initial segmentation of hyperechoic particle aggregation region: Input a complete anterior segment image (preferably a wide view) containing the anterior and posterior chambers, use the Otsu method to determine the global threshold and perform binary segmentation to obtain the hyperechoic foreground region.

[0044] Step 2 Morphological optimization and contour extraction: Perform morphological closing operation on the binary segmentation result, use 7×7 pixel circular structuring elements to connect adjacent hyperechoic foreground regions and smooth the boundaries; extract the boundary contour of each connected hyperechoic foreground region, preferably calculate the outer envelope contour (e.g. convex hull, i.e. the smallest convex polygon of the outer envelope of the region) for the connected regions to obtain a more stable boundary contour of the clustered region.

[0045] Step 3: Area Ratio Calculation: Count the total number of pixels within the convex hull of all high-echo particle aggregation regions, denoted as Area_aggregates_px; use the "Anterior Segment Scanning Region (Anterior Chamber + Posterior Chamber) ROI Mask" constructed in S2.1 to calculate Area_total_px; calculate... Aggregate_Ratio=(Area_aggregates_px / Area_total_px)*100%; Optionally, an Aggregate_Ratio > 5% may indicate the presence of clinically significant pyogenic aggregation.

[0046] S2.4 Determination of Tissue Infiltration Target ROI: Manually or semi-automatically select suspected infiltration areas (such as the anterior chamber angle crypts or the surface of the iris) in a narrow field of view. The selection width is fixed at 5 pixels (approximately 100μm) to cover the feature area and avoid excessive background.

[0047] Background ROI: Select a rectangular region with an area equal to that of the target ROI within the anechoic region of the vitreous cavity in the same frame image.

[0048] Calculation: Calculate the average gray level μ_target of the target ROI, the average gray level μ_bg of the background ROI, and the standard deviation σ_bg; calculate SNR=(μ_target-μ_bg) / σ_bg; when SNR>3.0, tissue infiltration is considered to be present. The threshold of 3.0 is determined based on a retrospective analysis of 50 infiltration-positive / negative cases (specificity>90%, sensitivity>85%).

[0049] S2.5 Postendothelial deposit density of cornea Step 1: Narrow-band ROI definition: Based on the corneal endothelial boundary L_upper, generate a band-shaped ROI with a constant width along L_upper towards the inside of the eye. The width is set to 15μm (equivalent to approximately 0.75 pixels according to a scale of 20μm / pixel; in practice, a continuous band-shaped ROI with a width of 1 pixel is used).

[0050] Step 2 Local Adaptive Threshold Segmentation: Calculate μ_local and σ_local for each local neighborhood (e.g., a 15×15 pixel window) within the narrowband ROI, and determine the threshold according to T_local=μ_local+k*σ_local, where k=1.5; pixels with gray values ​​higher than T_local are judged as deposition candidates.

[0051] Step 3 Noise Filtering: Perform morphological opening operation on the candidate binary image, and use 3×3 cross-shaped structuring elements to remove isolated noise points while preserving the true morphology of the deposited particles.

[0052] Step 4 Density Calculation: Count the number of connected components N_deposit after opening operation; convert the actual area of ​​the narrowband ROI A_band_mm² according to the pixel pitch (which can be approximated by "narrowband curve length × bandwidth"); calculate Deposit_Density=N_deposit / A_band_mm², in units of units / mm².

[0053] S3 Spatial Weighted Quantitative Assessment (see...) Figure 2 ) (1) Feature normalization and fusion: The continuous physical quantization features obtained in S2 are normalized according to a preset statistical range; for the wide-view image, the arithmetic mean of the 5 normalized features is taken to obtain the feature value P_center of the central region; for the i-th azimuth narrow-view image (i=1~8), its 5 normalized features are fused with the azimuth infiltration determination (0 / 1 or the corresponding SNR normalized value) to obtain the azimuth feature value P_i. Preferably, the normalization adopts truncated linear normalization based on statistical samples: for each feature, the P5 and P95 quantiles of the training / historical samples are taken as the lower / upper limits, truncated first and then mapped to [0,1] according to x′=(x−P5) / (P95−P5) to reduce the influence of outliers and ensure cross-batch comparability.

[0054] (2) Spatial weight matrix: Divide the anterior segment cross section into 8 orientations, corresponding to 8 narrow view images; construct a weight vector W=[w1,…,w8]. The lower orientation set S_inf includes at least the 6:00 orientation, and optionally further includes the adjacent 4:30 and / or 7:30 orientations; assign a weight coefficient of 1.5 to the orientations within S_inf, and assign a weight coefficient of 1.0 to the other orientations.

[0055] (3) ADCV calculation: Assume the harmonic coefficients α and β satisfy α + β = 1, calculate ADCV = 100×( α·P_center + β·( Σ(w_i·P_i) / Σw_i ) ), Where w_i is the weight coefficient of the i-th orientation. This calculation uses “Σw_i normalization + ×100 mapping” to make ADCV stably fall within the scale of 0 to 100, which is convenient for consistency with the S4 risk threshold.

[0056] S4 Structured Quantization Result Generation and Output The output module receives ADCV, various physical quantization feature values, and quality control information (final gain value, average linewidth W, acquisition sequence integrity), generates a structured report, and displays / exports it. (1) Core quantitative data: Planktonic particle density (particles / mm²), artifact intensity (contrast value), percentage of area of ​​hyperechoic particle aggregation region (%), tissue infiltration determination and corresponding SNR, corneal endothelial deposition density (particles / mm²), ADCV; (2) Automatic risk classification: ADCV is classified and indicated according to a preset threshold (e.g., ADCV<15 is low risk, 15-30 is medium risk, and ≥30 is high risk). The classification threshold can be determined by retrospective case data combined with clinical outcomes or expert classification labels. ROC analysis or Youden index is preferred to determine the threshold to achieve a classification effect that takes into account both sensitivity and specificity.

[0057] (3) Auxiliary control score: The physical quantity can be optionally mapped to a control score of 0 to 36 points for comparison with the traditional semi-quantitative system; (4) Visual annotation and historical comparison: The particle location, boundary of hyperechoic particle aggregation area, and infiltrated bright area are superimposed on the original image, and the S_inf orientation is highlighted; the ADCV trend over time can be output for multiple examinations of the same patient.

[0058] Example 2: Modular Implementation of the System like Figure 4As shown, the system includes: an acquisition control and monitoring module (performing S0 closed-loop quality control and acquisition sequence guidance), an image preprocessing module (performing S1a desensitization and S1b masking), a feature extraction module (performing S2.1 to S2.5 in parallel), a spatial weighted quantization module (storing the weight matrix W and α, β and performing S3 calculation), and an output module (generating a structured report and outputting / displaying it).

Claims

1. A UBM image quantitative analysis method for evaluating silicone oil emulsification, characterized in that, Includes the following steps: S0, Standardized Closed-Loop Control of Image Acquisition: Communicates with the Ultrasonic Biological Microscope (UBM) device to acquire acquisition parameters and / or real-time image quality indicators in real time; the acquisition parameters include at least a gain value G, and the quality indicators include at least a corneal echo width W; when the acquisition parameters and / or the quality indicators do not meet the preset standards, control commands and / or interactive prompts are generated to constrain the acquisition process, thereby obtaining a target image that meets the standardization requirements. S1. Image preprocessing: Desensitize the target image and mask the non-effective scanning areas to obtain the intraocular scanning area defined by the effective scanning area mask; wherein, the effective scanning area mask is generated and applied by thresholding the image and selecting the largest connected region. S2. Multi-dimensional acoustic feature extraction: Perform morphological operations, blob detection, threshold segmentation, texture analysis and / or signal-to-noise ratio calculation on the preprocessed image to extract continuous physical quantization features for quantizing the distribution of the medium; S3. Spatial weighted quantization evaluation: The continuous physical quantization features are normalized and fused to obtain the central region feature value P_center and the feature values ​​P_i of each direction; a spatial weight matrix is ​​introduced to assign weight coefficients w_i to different directions and the weighted acoustic distribution feature value ADCV is calculated; wherein P_center and P_i are feature values ​​obtained by fusing the normalized continuous physical quantization features; S4. Output Results: Output the ADCV and / or the structured quantization results associated with it.

2. The method according to claim 1, characterized in that, In step S0, determining whether the image quality meets the preset standard based on the acquired parameters specifically includes: (1) Read the gain value G (in dB) of the UBM device in real time and determine whether G is in the range of 79dB to 81dB; when G<79dB or G>81dB, generate a lock command to pause image freezing and saving, and generate an interactive prompt to guide the adjustment of G to the preset range. And / or, (2) For each frame of the acquired video stream, perform edge detection in the central region of the image to extract the corneal echo line; obtain the grayscale profile along the normal direction of the corneal echo line and fit it to calculate the full width at half maximum (FWHM) to obtain the corneal echo line width W (unit pixel); determine whether W is less than or equal to a 3-pixel threshold; when the average value of W for a consecutive preset number of frames exceeds the threshold, generate a probe adjustment prompt.

3. The method according to claim 1, characterized in that, The control instructions in step S0 include at least: disabling the freezing / saving of the current frame, sending a lock instruction to the device to pause freezing and saving, and unlocking and allowing saving when the acquisition parameters and / or the quality indicators recover to meet the preset standards.

4. The method according to claim 1, characterized in that, The target image includes one wide-view image and eight narrow-view images from different angles, with the angles including at least 1:30, 3:00, 4:30, 6:00, 7:30, 9:00, 10:30, and 12:

00.

5. The method according to claim 1, characterized in that, The continuous physical quantification features include at least one of the following: anterior chamber floating particle density, corneal endothelial multiple reflection artifact intensity, area ratio of hyperechoic particle aggregation region, tissue infiltration signal-to-noise ratio, and corneal endothelial deposition density; wherein, the area ratio of hyperechoic particle aggregation region is the area ratio of hyperechoic foreground connected region in the intraocular scanning region; the intraocular scanning region is the intraocular effective scanning region defined by the effective scanning region mask in step S1.

6. The method according to claim 1 or 5, characterized in that, The extraction of the anterior chamber airborne particle density includes: (1) Determine the region of interest (ROI) in the UBM image and obtain the gray matrix I(x,y) within the ROI; wherein, the upper boundary of the ROI includes the corneal endothelial boundary curve, and the lower boundary of the ROI preferably includes the anterior surface boundary curve of the iris, and when the upper edge of the anterior capsule of the lens can be reliably identified, the lower boundary can be further completed or corrected by the upper edge boundary curve of the anterior capsule of the lens; (2) Perform a morphological top-hat transformation on I(x,y) to obtain a grain-enhanced image I_th(x,y), where I_th(x,y) = I(x,y) - Open(I(x,y),B); Open(·,B) represents the morphological opening operation performed with structuring element B, where B is a circular structuring element; (3) Perform Laplacian-Gaussian LoG convolution on I_th(x,y) to obtain the response map R_σ(x,y)=∇²(G_σ*I_th)(x,y), and use a multi-scale detection strategy to calculate the response map for at least two different scale parameters σ respectively; (4) Select local extreme points that satisfy the preset gray level threshold in the response graph as candidate particle centers, and generate a set of candidate particle connected components on I_th(x,y) with the candidate particle centers as seeds. (5) Calculate the area A_k and perimeter P_k of each candidate particle's connected region, and calculate the roundness of the shape according to the roundness C_k=4πA_k / P_k²; and obtain the target particle set based on the preset gray level threshold and the roundness of the shape threshold. (6) Count the number of connected components N corresponding to the target particle set after screening; convert the ROI pixel area to the actual area A_mm² according to the pixel spacing label (0028,0030) in the DICOM header file; calculate the planktonic particle density Density=N / A_mm², where Density is in units of particles / square millimeter.

7. The method according to claim 6, characterized in that, The scale parameter σ of the multi-scale detection in step (3) includes at least 1.5 pixels, 2.5 pixels and 3.5 pixels, and the maximum value of the response at each scale is used as the final spot intensity of each pixel.

8. The method according to claim 1, characterized in that, The quantification of the intensity of the retroepithelial multiple reflection artifact includes: (1) Artifact analysis area determination: Based on the corneal endothelial boundary extracted from the corneal echo line in the UBM image, the area is offset by a preset physical distance along the intraocular direction and converted into pixel displacement according to the pixel spacing to obtain a strip-shaped area as the artifact analysis area. (2) Texture enhancement: Perform vertical Sobel gradient operator convolution on the grayscale image within the artifact analysis area to obtain the texture enhancement image; (3) Texture quantization: Construct a gray-level co-occurrence matrix (GLCM) in the artifact analysis area of ​​the texture enhancement map, set the pixel pair distance d and orientation angle θ, and calculate the contrast feature value according to Contrast=Σ(ij)²·P(i,j); (4) Output the contrast feature value as the Artifact_Strength quantification index.

9. The method according to claim 1, characterized in that, The spatial weighted quantitative assessment includes: (1) Normalize and fuse the continuous physical quantization features to obtain the central region feature value P_center and the feature values ​​P_i of each direction; (2) Construct a spatial weight matrix to divide the eyeball orientation into 8 orientations, the 8 orientations corresponding to the 8 narrow field images respectively, and assign a weight coefficient higher than that of other orientations to the lower orientation set S_inf, wherein S_inf includes at least the 6:00 orientation, and preferably further includes the adjacent 4:30 and / or 7:30 orientations; (3) Calculate the weighted acoustic distribution eigenvalue ADCV using the following formula: ADCV = 100×( α·P_center + β·( Σ(w_i·P_i) / Σw_i ) ) Where w_i is the weight coefficient of the i-th orientation, and α and β are preset harmonic coefficients with α+β=1.

10. A UBM image quantification analysis system for evaluating silicone oil emulsification, characterized in that, include: The acquisition control and monitoring module is used to perform the standardized closed-loop control of image acquisition in step S0 of claim 1; The image preprocessing module is used to perform the image desensitization and masking processing in step S1 of claim 1; The feature extraction module is used to perform the multi-dimensional acoustic feature extraction in step S2 of claim 1; A spatial weighted quantization module is used to perform the spatial weighted quantization evaluation and calculate ADCV in step S3 of claim 1. An output module is used to output the result of step S4 as described in claim 1.

Images