Iron death cell region automatic labeling method and system

Abstract

The present application relates to the technical field of image analysis, and particularly relates to a ferroptosis cell region automatic labeling method and system. The method comprises the following steps: acquiring multi-channel image data, wherein the multi-channel image data comprises ROS channel data, iron ion probe channel data and mitochondrial morphology channel data; performing region selection according to the multi-channel image data to obtain image region data; performing spatial offset alignment according to the image region data to obtain region alignment data; performing cross-channel correlation on the region alignment data to obtain cross-channel correlation data; performing confidence score according to the cross-channel correlation data to obtain region confidence data; and performing pseudo-label output according to the region confidence data to obtain region labeling data. The present application can automatically complete the recognition and labeling of ferroptosis cell regions based on the cross-modal feature cooperation of multi-channel images without manual labeling, thereby improving the data processing efficiency and the consistency of region recognition.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to an automatic annotation method and system for iron death cell regions. Background Technology

[0002] Ferroptosis is a novel programmed cell death mechanism closely associated with various diseases, often accompanied by typical cellular changes such as elevated reactive oxygen species (ROS) levels, iron accumulation, and mitochondrial structural damage. With the widespread application of high-throughput live-cell imaging technology, researchers can simultaneously acquire multimodal data such as ROS signals, iron ion distribution, and mitochondrial morphology through multi-channel fluorescence imaging, providing a data foundation for the identification and analysis of ferroptosis cells. However, current identification of ferroptosis regions mainly relies on manual observation of images from different channels and empirical annotation, which is not only inefficient and highly subjective but also difficult to guarantee consistency across samples. Especially when there are temporal or spatial shifts in signals from different channels, how to fuse asynchronous multi-channel information to identify ferroptosis-related regions has become one of the key challenges in this field. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes an automatic labeling method and system for iron death cell regions, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides an automatic labeling method for ferroptosis cell regions, including the following steps: Step S1: Acquire multi-channel image data, which includes ROS channel data, iron ion probe channel data, and mitochondrial morphology channel data; select regions based on the multi-channel image data to obtain image region data; Step S2: Perform spatial offset alignment based on the image region data to obtain region alignment data; Step S3: Perform cross-channel correlation on the region-aligned data to obtain cross-channel correlated data; Step S4: Calculate the confidence score based on the cross-channel correlation data to obtain regional confidence data; output pseudo-labels based on the regional confidence data to obtain regional annotation data.

[0005] This invention, by introducing joint analysis of ROS channels, iron ion probe channels, and mitochondrial morphology channels, can leverage the asynchronous cross-channel response of ferroptosis at the molecular and structural phenotype levels to achieve precise pre-screening control of regions. By setting spatial offset alignment and weak cross-channel consistency associations, the stability of region expression under conditions of multi-channel imaging misalignment and inconsistent response scales is improved. Through confidence quantification and pseudo-label generation, the reliability inconsistency problem caused by traditional reliance on manual annotation or hard threshold decision-making is solved, providing structured and hierarchical label support for unsupervised or weakly supervised training.

[0006] Preferably, acquiring multi-channel image data specifically involves: ROS channel data is acquired using the ROS imaging channel acquisition module; Data from the iron ion probe channel was acquired using the iron ion probe imaging module. Mitochondrial morphology channel data were acquired using a mitochondrial structure imaging module.

[0007] This invention achieves multidimensional imaging of different types of biological abnormalities within cells by separately setting up a ROS imaging channel, an iron ion probe imaging module, and a mitochondrial structure imaging module. The ROS channel can capture oxidative stress signals in the early stages of ferroptosis, the iron ion probe channel directly reflects the spatial distribution of abnormal iron ion accumulation, and the mitochondrial morphology channel can monitor key phenotypic changes in ferroptosis such as mitochondrial shrinkage and cristae structure destruction. By separately acquiring data from the three channels while maintaining consistent imaging coordinates, independent modeling and cross-channel comparisons can be performed on the data from different channels during analysis, improving the accuracy of identifying the evolutionary process of ferroptosis and the precision of regional localization.

[0008] Preferably, the region selection is specifically as follows: Preliminary regional screening is performed based on multi-channel image data to obtain preliminary regional screening data. Based on the initial regional screening data, cross-channel spatial neighborhood association is performed to obtain neighborhood association data; Weakly consistent region combination determination is performed on the neighboring related data to obtain image region data.

[0009] This invention introduces multi-channel response features in the initial screening stage of regions to achieve rapid local anomaly intensity localization and avoid redundant processing of large-area invalid regions. Through a cross-channel spatial neighborhood association strategy, instead of relying solely on pixel-level overlap, it utilizes the potential synergistic characteristics of spatial adjacency between multiple channels to enhance the fault tolerance of region retention and the information complementarity between channels. Based on a weak consistency region combination judgment strategy, a three-level judgment mechanism of "existence-proximity-scale rationality" is constructed to effectively eliminate pseudo-candidate regions driven by disordered channel responses, spatial structural conflicts, or random noise.

[0010] Preferably, the regional preliminary screening specifically includes: Cross-channel center point data is obtained by determining the center point across multiple channel image data. Intensity curves of channel spatial influence are constructed from cross-channel center point data to obtain intensity curve data; The diffusion radius of the channel is estimated based on the intensity curve data to obtain the diffusion radius data; Based on the diffusion radius data, the cross-channel diffusion radius relationship is determined, and the diffusion radius relationship data is obtained; Irrelevant diffusion regions are excluded from multi-channel image data based on diffusion radius relationship data to obtain preliminary regional screening data.

[0011] This invention avoids the bias problem of focusing on the intensity extreme value of a single channel by extracting the center point across channels, making the screening starting point more based on cross-modal expression; the constructed channel spatial influence intensity curve can systematically characterize the law of attenuation of abnormal signals of each channel with distance; by estimating the diffusion radius and determining its relationship, it can identify the "enclosed" or "proximity" spatial propagation structure between channels during ferrodeogenesis, and strengthen the ability to exclude non-hierarchical, non-diffusion structure response regions; through the multi-channel diffusion structure inconsistency exclusion strategy, it effectively reduces the risk of misidentification caused by imaging noise, local staining abnormalities or transient pseudo-activation.

[0012] Preferably, the determination of weakly consistent region combinations is specifically as follows: The minimum support across channels is determined for the neighborhood-related data to obtain primary regional unit data. Based on the spatial proximity consistency determination of the primary regional units, the secondary regional unit data is obtained; The consistency of regional scale relationships is determined by analyzing the secondary regional unit data to obtain the regional unit data. Image region data is obtained by hierarchically summarizing the primary region unit data, secondary region unit data, and region unit data.

[0013] In this invention, the minimum support determination across channels ensures that a region receives cooperative response support from at least two channels, effectively eliminating isolated noise from a single channel. The spatial proximity consistency determination avoids erroneous merging of spatially unrelated but numerically similar regions, strengthening the physical spatial coupling between candidate regions. Through the consistency determination of region scale relationships, candidate combinations with severely inconsistent areas and shapes between channels are excluded, enhancing the biological rationality and morphological consistency of the region structure. The organic integration of results at each level not only preserves regions with weak structural consistency but also addresses the issue of incomplete region overlap caused by differences in the response characteristics of different channels.

[0014] Preferably, step S2 specifically includes: Regional spatial features are generated based on image region data to obtain regional spatial feature data; Based on the regional spatial feature data, initial cross-channel regional matching is performed to obtain regional matching data; The matching offset is calculated on the region matching data to obtain the matching offset data; Based on the matching offset data, the image region data is spatially offset corrected to obtain preliminary alignment data; Fine-tuning of the initial alignment data by local spatial offset is performed to obtain the region alignment data.

[0015] This invention extracts spatial features of image regions (such as centroid coordinates, boundary contours, area and shape) to provide a stable geometric description basis for inter-channel region matching; the initial matching and offset calculation module can quickly identify the overall offset trend between multi-channel regions and complete the global correction of the main misalignment; through fine adjustment of local spatial offset, it effectively addresses the non-rigid offset problem caused by sample drift, micro-imaging distortion or optical path differences, and achieves flexible registration at the region level.

[0016] Preferably, step S3 specifically includes: Cross-channel associated region filtering is performed on the region-aligned data to obtain cross-channel associated region data; Regional spatial consistency data is obtained by performing regional spatial consistency calculations based on cross-channel associated regions; Based on the regional spatial consistency data, the consistency of channel feature association is determined, and the channel feature association data is obtained; Multi-region association groups are constructed based on channel feature association data to obtain cross-channel association data.

[0017] This invention uses cross-channel associated region screening to identify candidate region pairs with synergistic change trends or spatial interaction potential in aligned regions, laying the foundation for multimodal aggregation. The system performs quantitative calculations based on the spatial consistency between regions, clarifying the relative offset, overlap, and structural coordination between regions in each channel. By judging the consistency of channel feature associations and combining the matching of features such as response intensity and dynamic trends between multiple channels, pseudo-associated regions caused by accidental overlap are eliminated. By constructing multi-region association groups, the combination and merging of multi-channel regions at the response mechanism level is realized, effectively improving the ability to capture and express the asynchronous response phenomenon of iron death.

[0018] Preferably, the confidence score is as follows: Based on cross-channel correlation data, regional alignment stability assessment, cross-channel support statistics, and channel change analysis are performed to obtain regional alignment stability data, cross-channel support data, and channel consistency data, respectively. Confidence scores are calculated based on regional alignment stability data, cross-channel support data, and channel consistency data to obtain regional confidence scores.

[0019] In this invention, the region alignment stability assessment measures the positional consistency and deformation degree of each region across multiple channels from a spatial registration perspective, reflecting its geometric rationality as a response region to the same biological event. Cross-channel support statistics are used to identify whether a region receives support from multi-channel collaborative responses, strengthening the control over the dependence on multimodal consistent expression. Through channel change analysis, the response intensity, change magnitude, and dynamic trend of a region in each channel are comprehensively analyzed to ensure that the scoring system is sensitive to the evolutionary characteristics of biological signals. By jointly modeling and calculating the region confidence score using the above three types of heterogeneity indicators, not only are single-factor misjudgments or human threshold dependence problems avoided, but the reliability and interpretability of automatic annotation results are also improved.

[0020] Preferably, the pseudo-label output is as follows: Pseudo-label regions are filtered based on regional confidence data to obtain pseudo-label region data; Perform region verification based on pseudo-label region data to obtain pseudo-label verification data; Based on the pseudo-label verification data, the pseudo-label category and labeling format are determined to obtain pseudo-label data; Region annotations are generated based on pseudo-label data to obtain region annotation data.

[0021] This invention effectively filters low-confidence regions by screening regional confidence data, ensuring that only candidate regions with stable structures and clear responses are included in the pseudo-label generation range. A regional verification mechanism is set up to reconfirm the spatial consistency and channel feature rationality of candidate pseudo-label regions, improving the reliability and semantic accuracy of annotation quality. By classifying pseudo-label categories and determining annotation formats for regions with different confidence levels, the system has the ability to flexibly adapt to annotations of multiple categories and multiple confidence intervals. The generated regional annotation data can be used as weakly supervised labels for model training, and can also be used to support visualization analysis and expert review, reducing the cost of manual annotation.

[0022] Preferably, this application also provides an automatic labeling system for ferroptosis cell regions, used to perform the automatic labeling method for ferroptosis cell regions as described above, the automatic labeling system for ferroptosis cell regions comprising: The region selection module is used to acquire multi-channel image data, which includes ROS channel data, iron ion probe channel data, and mitochondrial morphology channel data; and to select regions based on the multi-channel image data to obtain image region data. The spatial offset alignment module is used to perform spatial offset alignment based on image region data to obtain region alignment data. The cross-channel association module is used to perform cross-channel association on region-aligned data to obtain cross-channel associated data; The region labeling module is used to score confidence based on cross-channel correlation data to obtain region confidence data; and to output pseudo-labels based on the region confidence data to obtain region labeling data.

[0023] The beneficial effects of this invention are as follows: By jointly acquiring image data covering multi-dimensional features such as oxidative stress, metal ion enrichment, and mitochondrial morphology changes through ROS channels, iron ion probe channels, and mitochondrial morphology channels, a complete pathological expression basis is provided for region screening and modeling; a cross-channel diffusion radius relationship determination mechanism is introduced in the region selection stage to effectively eliminate pseudo-response regions without biological structural consistency; in the spatial offset alignment stage, the systematic error and non-rigid deformation problems existing in multi-channel imaging are solved by combining region-level initial matching with local fine adjustment, improving the geometric stability of subsequent region matching; a joint modeling mechanism of structural consistency and response consistency is further introduced in the cross-channel association stage to realize region fusion under asynchronous expression of different channels; a multi-factor credibility evaluation system is constructed through confidence scoring to quantify the spatial stability, channel support, and response consistency of regions, providing a grading basis for the final pseudo-label generation; the final pseudo-label output step combines structural verification and category adaptation to realize automatic annotation generation for training, effectively reducing the dependence on manual annotation and improving the interpretability and scalability of annotation quality. Attached Figure Description

[0024] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings: Figure 1 A flowchart illustrating the steps of an automatic labeling method for iron-dead cell regions according to an embodiment is shown. Figure 2 A flowchart illustrating the steps of a multi-channel image data acquisition method according to an embodiment is shown. Figure 3 A flowchart illustrating the steps of a spatial offset alignment method according to one embodiment is shown. Figure 4 A flowchart illustrating the steps of a cross-channel association method according to one embodiment is shown; Figure 5 A flowchart illustrating the steps of a confidence scoring method according to an embodiment is shown. Detailed Implementation

[0025] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0027] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] The system is used for automatic region labeling of ferroptosis-related images of in vitro cultured cells. The system acquires three-channel fluorescence images from the same sample field of view, with a resolution of 1024×1024 pixels and a pixel size of 0.2 μm / pixel. ROS channel data was acquired using the DCFDA probe, iron ion probe channel data using the FerroOrange probe, and mitochondrial morphology channel data using the MitoTracker Deep Red FM. The system first performs background subtraction, flattening correction, crosstalk correction, and intensity normalization on the three-channel images, and then identifies candidate response centers in each channel. For any candidate center, the system calculates its diffusion radius in the ROS, iron ion probe, and mitochondrial morphology channels. If the ratio of the diffusion radii of at least two of the three channels is not greater than 2.0, and the centroid distance between the corresponding regions is not greater than 20 pixels, then this region is retained as image data; otherwise, it is excluded as an irrelevant diffusion region.

[0029] During spatial offset alignment, the system uses the ROS channel as a reference channel to perform centroid matching on candidate regions in the iron ion probe channel and mitochondrial morphology channel. The system calculates the centroid offset of each matched region and excludes matched pairs with a centroid distance greater than 20 pixels, an area difference exceeding ±30%, or an intersection-union ratio (IU) below 0.10 after translation. Subsequently, the system determines the overall spatial offset vector based on the median of the x and y offsets of the remaining matched pairs. For example, in a certain field of view, the overall offset of the iron ion probe channel relative to the ROS channel is (4, -2) pixels, and the overall offset of the mitochondrial morphology channel relative to the ROS channel is (-3, 3) pixels. Based on this, the system performs uniform translation correction on the target channel regions; then, local fine-tuning is performed within a 30×30 pixel local window, reducing the average centroid offset distance across channel regions from 8.6 pixels before correction to 2.4 pixels, and increasing the IU from 0.18 to 0.42, thus obtaining the region alignment data.

[0030] During cross-channel association, the system performs association judgment on aligned candidate regions. If two regions from different channels satisfy at least two of the following conditions: centroid distance not greater than 20 pixels, region intersection-union ratio not less than 0.10, minimum boundary distance not greater than 5 pixels, or relative response intensity not less than 20%, then a cross-channel association edge is established. The system determines the set of connected regions containing at least two regions from different channels as the cross-channel association group. For a certain cell region, if the average response intensity of the ROS channel is 1.8 times the average background intensity, the average response intensity of the iron ion probe channel is 1.7 times the average background intensity, and the mitochondrial morphology channel shows fragmented mitochondrial outlines and decreased edge continuity, and these three form the same cross-channel association group in space, then the system outputs this region as a suspected ferroptosis candidate region as cross-channel association data.

[0031] During the confidence scoring process, the system calculates region alignment stability, cross-channel support, and channel consistency. If the centroid offset distance of the candidate region is no greater than 5 pixels, the average nearest neighbor distance at the boundary is no greater than 10 pixels, and the region intersection-union ratio is no less than 0.30, then the region alignment stability is assigned a value of 0.9. If all three channels have effective responses, i.e., the average response intensity is no less than the average background intensity plus twice the background standard deviation, or no less than 1.5 times the median intensity of the corresponding channel's overall image, and there are no fewer than 30 effective response pixels, with the effective response pixel ratio being no less than 60%, then the cross-channel support is assigned a value of 0.9. If the response enhancement ratio of at least two channels is no less than 20%, and the intensity curves show the same diffusion or attenuation direction in at least three consecutive radius layers, then the channel consistency is assigned a value of 0.9. The system calculates the arithmetic mean of the three scores to obtain the confidence score for the region. If a region has a region alignment stability score of 0.9, a cross-channel support score of 0.9, and a channel consistency score of 0.6, then its confidence score is 0.8, and the system marks it as a medium-to-high confidence suspected iron death region; if all three are 0.9, then the confidence score is 0.9, and the system marks it as a high confidence iron death region.

[0032] Through the above processing, in a single multi-channel image containing approximately 300 cell regions, the system can automatically generate pixel-level pseudo-label masks, region bounding boxes, region category numbers, and confidence scores. Compared to manually observing and delineating regions channel by channel, this method reduces the region labeling time for a single image from approximately 20 minutes to less than 2 minutes. Simultaneously, by introducing spatial offset alignment and cross-channel correlation judgment, single-channel noise bright spots, unevenly stained regions, and non-co-localized fluorescence regions are effectively eliminated, reducing low-confidence mislabeled regions. Therefore, this embodiment can obtain region-labeled data with spatial consistency, cross-channel response support, and confidence levels without requiring manual region-by-region labeling, providing a structured data foundation for ferroptosis image sample screening, weakly supervised model training, and expert review.

[0033] Please see Figures 1 to 5 This application provides an automatic labeling method for ferroptosis cell regions, comprising the following steps: Step S1: Acquire multi-channel image data, which includes ROS channel data, iron ion probe channel data, and mitochondrial morphology channel data; select regions based on the multi-channel image data to obtain image region data; In one embodiment, intracellular reactive oxygen species (ROS) were labeled with DCFDA fluorescent dye at a final concentration of 10 μM and incubated at 37°C for 30 minutes. During imaging, the excitation wavelength was set to 488 nm, and the fluorescence emission acquisition band was from 500 nm to 550 nm, resulting in a green fluorescence image, which was used as the ROS channel image. FerroOrange probes were used to label intracellular ferrous ions (Fe2+). 2+ Fluorescent labeling was performed with a final probe concentration of 1 μM, and incubation was carried out at 37°C for 30 minutes. For imaging, the excitation wavelength was set to 561 nm, and the fluorescence emission acquisition band was 570 nm to 620 nm, yielding orange-red fluorescence images, which were used as iron ion channel images / iron ion probe channel data. Mitochondrial structures were labeled using MitoTracker Deep Red FM with a final dye concentration of 100 nM to 200 nM, and incubation was carried out at 37°C for 20 to 30 minutes. For imaging, the excitation wavelength was set to 633 nm or 640 nm, and the fluorescence emission acquisition band was 660 nm to 720 nm, yielding far-red fluorescence images, which were used as mitochondrial morphology channel images.

[0034] For the acquired three-channel images, the system first performs background intensity normalization and noise suppression on each frame, and then performs local response analysis using a sliding window to identify potential abnormal response centers where the response intensity / pixel value is higher than that of neighboring regions. Specifically, the system scans the entire image using a sliding window (e.g., 20×20 pixels), performs local statistical analysis within each window, and calculates the average intensity difference or Z-score between the center pixel and its neighboring pixels. When the intensity of the center pixel is higher than the neighborhood average and exceeds a set response threshold (e.g., higher than the mean + 2.5 × standard deviation), the center location is identified as an abnormal response center. Based on the spatial distribution relationship between different channels, the system evaluates the spatial diffusion range of each response region and selects regions with spatial proximity or nesting features in at least two channels as image regions. In other words, the system evaluates cross-channel spatial relationships based on the centroid location, boundary contour, and spatial overlap of response regions in different channels. For each identified response region / image region in a channel, the system calculates its spatial proximity (e.g., centroid distance less than a set threshold, such as 15 pixels), overlap area ratio (e.g., intersection area to union area ratio greater than 0.1), or whether boundary nesting exists (e.g., one channel region is partially enclosed within another channel region, with an enclosing ratio of a preset threshold, such as 10% or more). When the response regions of any two channels satisfy one of the above spatial proximity or nesting relationships, the system merges them into the same response region / image region. The system outputs the corresponding spatial location of the identified image region in the three channels and its multi-channel response matrix.

[0035] Step S2: Perform spatial offset alignment based on the image region data to obtain region alignment data; In one embodiment, to address the issue of regional response position shifts caused by differences in the imaging process in multi-channel images, the system extracts spatial features from the image regions extracted from each channel. This includes obtaining the centroid position, area, and boundary contour information of each region, and constructing a spatial feature vector for the region accordingly. Using the ROS channel as a reference coordinate system, initial registration operations are performed on the regions in the iron ion channel and the mitochondrial morphology channel, respectively. The system performs preliminary matching between the regions of the reference channel and the target channel using a centroid nearest neighbor search method, and combines this with an anomaly removal strategy based on consistency judgment to exclude region pairs with large matching errors, thereby obtaining a more accurate set of initial matching pairs. After obtaining the preliminary matching pairs, the system calculates the centroid displacement difference between each matching pair, and estimates the overall spatial offset direction and magnitude of the target channel relative to the reference channel. Then, global translation correction is performed on the target channel image to complete the first stage of alignment processing. For regions with local nonlinear deformation, the system performs local affine transformation within the neighborhood of each candidate region. The alignment error is optimized and adjusted based on the consistency of region features within the neighborhood, thus achieving local non-rigid registration. For example, the system uses the set of boundary points of each candidate region in the reference and target channels as input. Based on these boundary point coordinate pairs, the system estimates a set of affine transformation parameters. This affine transformation consists of a two-dimensional linear transformation matrix and a displacement vector. The former represents linear transformation behaviors such as local scale changes, rotations, or tilts, while the latter represents the overall spatial translation. By minimizing the overall matching error between the transformed coordinates of the target channel boundary points and the reference channel boundary points, the system solves for the optimal linear transformation matrix and offset. The system applies the above affine transformation to all boundary points of the corresponding region in the target channel, achieving higher spatial consistency with the region contour in the reference channel. The system outputs region alignment data after global correction and local fine-tuning.

[0036] Step S3: Perform cross-channel correlation on the region-aligned data to obtain cross-channel correlated data; In one embodiment, for spatially aligned image region data, the system constructs cross-channel candidate associated region pairs to identify regional responses reflecting the same biological event in different channels. The system calculates two types of indices between each candidate region pair: the first is a spatial consistency index, primarily used to evaluate the geometric matching degree of regions in different channels, including the degree of region overlap, the normalized distance of the centroid position, and the degree of shape difference between boundary contours; the second is a feature consistency index, used to compare the image feature performance of regions in different channels, mainly including the response intensity ratio of reactive oxygen species to iron ion signals, the degree of symmetry of region morphology, and the consistency score of edge texture features. The system sets a set of multi-factor weak consistency judgment thresholds for tolerant identification of potential associated regions. As long as any two channels have region pairs that simultaneously satisfy spatial overlap or proximity and have similar feature expression directions, they can be identified as valid cross-channel associated regions. All region pairs that meet this judgment condition will be combined into multi-region association groups / cross-channel associated data to represent candidate image region units for the same potential ferroptosis event.

[0037] Step S4: Calculate the confidence score based on the cross-channel correlation data to obtain regional confidence data; output pseudo-labels based on the regional confidence data to obtain regional annotation data.

[0038] In one embodiment, the system normalizes or transforms the above three types of indicators into several confidence levels (e.g., high, medium, low), assigning a specific value to each level, such as 0.9, 0.6, and 0.3. The system then averages and sums the specific values ​​corresponding to different indicators to obtain the region confidence data. For region alignment stability, if the centroid offset distance between the candidate region and the target channel is no more than 5 pixels, the average nearest neighbor distance at the boundary is no more than 10 pixels, and the region intersection-union ratio is no less than 0.30, then the region alignment stability is determined to be of the high level and assigned a value of 0.9; if the centroid offset distance is greater than 5 pixels but not greater than 10 pixels, and the region intersection-union ratio is no less than 0.20, then it is determined to be of the medium level and assigned a value of 0.6; if the centroid offset distance is greater than 10 pixels, or the region intersection-union ratio is less than 0.20, then it is determined to be of the low level and assigned a value of 0.3. For cross-channel support, if there are valid responses in the ROS channel, iron ion probe channel, and mitochondrial morphology channel, the cross-channel support is classified as high-level and assigned a value of 0.9; if there are valid responses in two of the channels, it is classified as medium-level and assigned a value of 0.6; if only one channel has a valid response, or if there is abnormal highlighting in a single channel while other channels do not meet the valid response standard, it is classified as low-level and assigned a value of 0.3. A valid response is defined as an average response intensity of the candidate region in the corresponding channel that is not less than the average background intensity plus twice the background standard deviation, or not less than 1.5 times the median intensity of the entire image in that channel, with a valid response area of ​​not less than 30 pixels and a valid response pixel ratio of not less than 60%. For channel consistency, if the average response intensity of at least two channels is at least 20% higher than the background region, and the intensity curves show the same direction of enhancement, diffusion, or attenuation in at least three consecutive radius layers, and the correlation coefficient between the changing directions of the channels is at least 0.6, and there are no channels with opposite response directions, then the channel consistency is judged as high-level and assigned a value of 0.9; if only two channels meet the condition of consistent response trends, and the third channel does not show an opposite changing direction, then it is judged as medium-level and assigned a value of 0.6; if only one channel shows a significant response trend, or any channel shows a changing direction opposite to other channels, then it is judged as low-level and assigned a value of 0.3. The system determines the regional confidence level based on the regional confidence score C. When C is not lower than 0.8 and there are no low-level indicators among the three indicators, the system marks the region as a high-confidence region; when C is not lower than 0.5 but lower than 0.8, or when one of the three indicators is low-level but the other two are high-level, the system marks the region as a medium-confidence region; when C is lower than 0.5, or when two or more of the three indicators are low-level, the system marks the region as a low-confidence region. Based on these rules, the system determines the confidence level of each region and outputs the corresponding region confidence data.

[0039] The system performs pseudo-label screening. In static screening mode, regions with a confidence score of at least 0.85 are designated as high-confidence pseudo-label regions. In dynamic screening mode, the system selects the top 10% of regions as pseudo-label regions. Regions with a confidence score between 0.65 and 0.85 can be considered as candidate regions for verification instead of being directly output as positive sample pseudo-labels. Local consistency verification is performed to check whether the region has a response pattern change direction opposite to that in its temporal or spatial neighboring regions, with a change amplitude of at least 30% of the intensity of the previous frame, or a centroid displacement exceeding 10 pixels, or a region area change ratio exceeding 50%. If any of these exist, the region is marked as unstable and excluded from the pseudo-label data. Stable regions that pass the verification are labeled with pseudo-labels, with the pseudo-label category set as "suspected iron death region," and corresponding pixel-level mask data is generated for training or verification. The final pseudo-label results are organized into a structured annotation data format, including image number, region location, category number, confidence score, and mask data.

[0040] Preferably, acquiring multi-channel image data specifically involves: Step S11: Acquire ROS channel data through the ROS imaging channel acquisition module; In one embodiment, the system acquires images of intracellular reactive oxygen species (ROS) signals using a ROS imaging channel acquisition module on a confocal fluorescence microscope (e.g., a Zeiss LSM880). In the experiment, DCFDA (dichlorofluorescein diacetate) was used as the ROS fluorescent probe. Cell samples were pretreated with a final dye concentration of 10 μM and incubated at 37°C for 30 minutes to ensure sufficient labeling of ROS signals. During imaging, the system set the excitation wavelength to 488 nm to excite the green fluorescence signal formed after DCFDA labeling. The fluorescence emission acquisition band was set to 500 nm to 550 nm, acquiring green fluorescence images as ROS channel images. The image acquisition resolution could be set to 1024 × 1024, the pixel size to 0.2 μm, the exposure time to 150 ms, and the output image format to 16-bit grayscale. The system encapsulates the pixel intensity matrix, image number, acquisition time, and imaging parameters of the acquired ROS channel images to obtain ROS channel data. The ROS channel data is used to characterize the spatial distribution features of intracellular oxidative stress response and is suitable for identifying regions of reactive oxygen species overexpression.

[0041] Step S12: Acquire iron ion probe channel data through the iron ion probe imaging module; In one embodiment, the system acquires images of intracellular ferrous ion response using an iron ion probe imaging module. In the experiment, the iron-specific fluorescent probe FerroOrange was used to detect intracellular ferrous ions (Fe...2+ Fluorescent labeling was performed, and the cells to be tested were incubated with FerroOrange at a final concentration of 1 μM at 37°C for 30 minutes to ensure that the probe fully entered the cells and reacted with Fe. 2+ After incubation, the system washes the cell samples with buffer to reduce background fluorescence interference caused by the free probes, and then immediately performs imaging. Imaging is performed using a laser scanning confocal system (such as Nikon A1+ or Olympus FV3000). The system is set with an excitation wavelength of 561 nm and a fluorescence emission acquisition band of 570 nm to 620 nm to capture the orange-red fluorescence signal generated by FerroOrange, thus acquiring iron ion channel images. The image acquisition depth can be set to 16-bit grayscale format, the single-frame exposure time can be set to 200 milliseconds, and averaging of every two image frames can be used to reduce random noise. The system performs grayscale normalization on the iron ion channel images to obtain iron ion probe channel data. The iron ion probe channel data is used to characterize intracellular Fe 2+ The spatial distribution of anomalous enrichment provides image evidence for subsequent determination of whether candidate regions have iron death-related metal ion accumulation characteristics.

[0042] Step S13: Acquire mitochondrial morphology channel data using the mitochondrial structure imaging module.

[0043] In one embodiment, the system acquires mitochondrial morphology channel data through a mitochondrial structure imaging module. To avoid channel cross-coloration caused by both mitochondrial and ROS channels using 488nm excitation and green fluorescence emission, the system uses MitoTracker Deep Red FM as a mitochondrial-specific fluorescent dye to label mitochondrial structures. In the experiment, the final concentration of MitoTracker Deep Red FM was set to 100nM to 200nM, and cell samples were incubated at 37°C for 20 to 30 minutes to enrich the dye in the mitochondrial structural regions. During imaging, the system set the excitation wavelength to 633nm or 640nm and the fluorescence emission acquisition band to 660nm to 720nm, acquiring far-red fluorescence images as mitochondrial morphology channel images. The system extracts the pixel matrix, mitochondrial edge response, local texture distribution, and morphological parameters of the corresponding images and performs structured encapsulation to obtain mitochondrial morphology channel data. This mitochondrial morphology channel data is used to characterize changes in mitochondrial structural integrity during ferroptosis.

[0044] Preferably, the region selection is specifically as follows: Preliminary regional screening is performed based on multi-channel image data to obtain preliminary regional screening data. In one embodiment, local peak detection is performed independently in each channel image. Specifically, this includes Gaussian filtering the image to reduce background interference and extracting local extrema points as the initial response point set for that channel. The system constructs a series of annular regions with increasing radii centered on each initial response point to analyze the spatial diffusion characteristics of the response. Within each annular region, the system calculates the corresponding average response value and plots the trend of response intensity changing with distance. When the average response value within a certain radius drops below a specific percentage of the maximum response value within the channel (e.g., below 30%), the system records this distance as the response diffusion radius of that channel. The system compares the diffusion radii of the three channels with the same center point as the origin. If the difference between the diffusion radii of any two channels is no greater than 10 pixels, or their ratio is between 0.5 and 2.0, or they exhibit specific nested structure characteristics (e.g., the diffusion range of the ROS channel is greater than that of the iron ion channel, and the iron ion channel is greater than that of the mitochondrial channel), then the center point is considered to have consistent abnormal response characteristics in the multi-channel image and can be used as the region center. The system constructs circular regions based on the center of each region and its maximum diffusion radius in the three channels, and outputs the initial screening data for each region.

[0045] Based on the initial regional screening data, cross-channel spatial neighborhood association is performed to obtain neighborhood association data; In one embodiment, the system performs cross-channel spatial neighborhood association processing on the preliminary screening region data obtained in the aforementioned steps. Each preliminary screening region includes center coordinates, diffusion radius, and channel information, serving as a region representation unit. The system performs pairwise comparisons of region sets / region representation units in any two channels, calculating the spatial distance between the centers of each region. When the center distance between two regions is less than or equal to the sum of their diffusion radii plus a preset spatial tolerance (e.g., 5 pixels), it is considered that the two regions have effective spatial overlap or proximity, possessing neighborhood association characteristics. For all region pairs that meet the above conditions, the system establishes cross-channel neighborhood association edges to indicate that two regions in different channels have a spatial overlap or adjacency relationship. The set of these association edges constitutes the neighborhood association data, where each data item represents a group of regions with spatial consistency in different channels, all pointing to the same potential iron death-related event.

[0046] Weakly consistent region combination determination is performed on the neighboring related data to obtain image region data.

[0047] In one embodiment, the system performs channel support judgment. The system checks whether each group of neighboring regions contains regions from at least two different imaging channels. If this basic cross-channel coverage requirement is met, the system proceeds to the next step of analysis. The system performs spatial consistency judgment. The system evaluates the spatial relationship between the regions in the group, calculating the distance between the center points of each region. If all center distances are less than a set adjacency threshold (e.g., 10 pixels), the group of regions is considered to have sufficient spatial proximity. The system performs diffusion scale judgment. The system compares the diffusion radius of the regions in each channel. If the diffusion scales between channels are relatively close (e.g., the ratio between the maximum and minimum diffusion range does not exceed a specific upper limit, such as 2.0), or if a region in one channel can cover or be contained within a region in another channel, it indicates a certain degree of coordination in spatial response patterns. When all three conditions are met, the system considers the region group as a weakly consistent cross-channel response unit and merges all regions within the group. This can be done using geometric union or by constructing a circumcircle to generate a unified image region unit. All regions that meet the weak consistency combination conditions are summarized and output as image region data.

[0048] Preferably, the regional preliminary screening specifically includes: Cross-channel center point data is obtained by determining the center point across multiple channel image data. In one embodiment, the system selects points with response values ​​higher than the average value of surrounding pixels within a set sliding window range, as the initial set of center points for that channel. The system performs spatial aggregation processing on the initial center points extracted from the three channels. When the spatial distance between two or more center points from different channels is less than a preset spatial overlap threshold (e.g., 10 pixels), the system merges these center points into a unified cross-channel center point and records its corresponding channel source information. The system outputs all aggregated cross-channel center point data, with each record including the spatial coordinates of the center point and the supported imaging channel type.

[0049] Intensity curves of channel spatial influence are constructed from cross-channel center point data to obtain intensity curve data; In one embodiment, the system constructs multiple concentric ring regions with increasing radii on each channel image, using the center point as the center, for example, gradually increasing from 1 pixel to 30 pixels, forming a set of continuous spatial region bands. Within each ring region, the system calculates the average response intensity of the channel image, thereby establishing a correspondence between spatial distance and signal response. By performing the above calculations for each channel separately, the system can obtain the spatial diffusion characteristics of the center point in different channels, forming an influence intensity curve describing the trend of signal intensity variation with spatial distance. The system uniformly organizes the intensity curve data corresponding to each cross-channel center point in each channel, collectively referring to it as intensity curve data.

[0050] The diffusion radius of the channel is estimated based on the intensity curve data to obtain the diffusion radius data; In one embodiment, the system normalizes the intensity curves of each channel to unify the numerical scale between different channels and avoid interference caused by differences in image brightness or response amplitude. Starting from the center of each curve, the system searches outwards along the radius, monitoring the trend of curve intensity changes in real time. When the intensity value first drops below a preset percentage of the peak intensity of the channel curve (e.g., to 30% or 40% of the maximum value within a preset radius), the system defines the corresponding spatial distance as the diffusion radius of the channel, characterizing the effective propagation range of the signal at the center point within that channel. If, within the set maximum search radius, there is no consecutive decreasing trend in at least three radius layers, and the cumulative decrease is not less than 20% of the peak intensity, the system treats the maximum search radius as the upper limit of the diffusion radius of the channel. Using this method, the system can generate a set of diffusion radius data for each cross-channel center point, corresponding to the ROS channel, iron ion probe channel, and mitochondrial morphology channel, respectively.

[0051] Based on the diffusion radius data, the cross-channel diffusion radius relationship is determined, and the diffusion radius relationship data is obtained; In one embodiment, the system determines whether the multi-channel diffusion radius satisfies any of the following types of relationships based on preset rules: First, whether it exhibits stable progressive or nested characteristics, for example, the diffusion radius of the ROS channel is greater than that of the iron ion channel, which in turn is greater than that of the mitochondrial morphology channel, forming an enclosing structure from the outside in; Second, whether the difference in diffusion radius between channels is within a reasonable range, without obvious abnormal proportional deviation or excessive separation, reflecting the synergy of signal propagation characteristics. When any of the above relationship patterns are met, the system considers the multi-channel diffusion characteristics of the center point to have spatial diffusion consistency, records it as a valid diffusion radius relationship, and outputs the corresponding diffusion radius relationship data.

[0052] Irrelevant diffusion regions are excluded from multi-channel image data based on diffusion radius relationship data to obtain preliminary regional screening data.

[0053] In one embodiment, the system performs validity screening on center points and their associated regions in multi-channel images based on diffusion radius relationship data. For center points determined not to meet the preset diffusion structure relationship rules, the system considers them to lack consistent or cooperative diffusion characteristics across different channels, belonging to non-specific responses or noise interference. Therefore, these center points and their corresponding spatial neighborhoods are marked as irrelevant diffusion regions and eliminated. For center points whose diffusion radius relationships are deemed valid, the system retains them and, using this center point as the center, combines its diffusion radius in each channel, selecting the maximum value as the coverage area to construct a circular candidate region. All candidate regions that pass the diffusion relationship consistency screening are uniformly combined into the initial region screening data.

[0054] Preferably, the determination of weakly consistent region combinations is specifically as follows: The minimum support across channels is determined for the neighborhood-related data to obtain primary regional unit data. In one embodiment, the system takes a pre-constructed cross-channel neighborhood association dataset as input, where each neighborhood association record represents a spatial adjacency group composed of candidate regions from different channels, including regional information from ROS channels, iron ion channels, and mitochondrial morphology channels. For each neighborhood association group, the system sequentially evaluates the validity of the response regions of each channel. The evaluation criteria include two aspects: first, whether the average response intensity of the region in the corresponding channel image is higher than a certain multiple (e.g., 1.2 times) of the overall median value of that channel; and second, whether the spatial area of ​​the region reaches a preset minimum threshold (e.g., an area of ​​not less than 30 pixels). When at least two regions from different channels in a neighborhood group simultaneously meet the above validity conditions, the system considers the neighborhood combination to have passed the minimum support determination. The valid regions contained in such neighborhood groups are integrated into a primary region unit, i.e., primary region unit data.

[0055] Based on the spatial proximity consistency determination of the primary regional units, the secondary regional unit data is obtained; In one embodiment, the system extracts the centroid coordinates of each channel region in the unit and calculates the spatial distance between any two regions in the channels. When the centroid distance between all regions in the primary unit is less than a set spatial proximity threshold (e.g., 20 pixels), or there is a significant spatial overlap (e.g., regions overlap to a certain extent, with an overlap rate exceeding a set threshold), the system considers the region unit to have consistency and synergy in spatial layout. For primary region units that pass the above spatial proximity consistency determination, the system identifies them as secondary region unit data.

[0056] The consistency of regional scale relationships is determined by analyzing the secondary regional unit data to obtain the regional unit data. In one embodiment, the system calculates the area of ​​each channel region, i.e., the corresponding number of pixels, and extracts the maximum and minimum area values ​​as comparison criteria. A region unit is considered to have scale consistency if it meets any of the following conditions: First, the area difference between the channel regions is small, for example, the ratio between the maximum and minimum areas does not exceed a set threshold (e.g., 2 times); second, there is a clear nesting or containment relationship between regions, for example, one region completely surrounds another region, and the boundary deviation is within an allowable range (e.g., within 3 pixels); third, the aspect ratios of each region are similar, and the difference between aspect ratios is controlled within a certain range (e.g., within ±0.5), indicating a certain consistency or homology in morphological structure. Sub-region units that pass the above scale consistency determination are recorded as region unit data by the system.

[0057] Image region data is obtained by hierarchically summarizing the primary region unit data, secondary region unit data, and region unit data.

[0058] In one embodiment, the system uses region unit data that has passed the scale relationship consistency judgment as the output result. Combined with relevant information from the previous primary support judgment and spatial proximity consistency judgment processes, it performs hierarchical induction processing to construct image region data with a unified format. While preserving the region structure, the system also records metadata for each region unit during the analysis process, including which imaging channels initially provided support, the spatial offset distance between channel regions, and the area differences or nesting relationships between regions. During the induction process, the system uniformly represents the position of each region unit, either by calculating the bounding box from the set of corresponding regions in multiple channels or by using the average of multiple centroid coordinates to represent the region center. Each region is assigned standardized attribute labels, such as the number of supporting channels and cross-channel response consistency score. The image region data output by the system is organized in a structured format, including fields such as region number, spatial location, list of supporting channels, and consistency score.

[0059] Preferably, step S2 specifically includes: Step S21: Generate regional spatial features based on image region data to obtain regional spatial feature data; In one embodiment, the system extracts the centroid coordinates of each region to represent its spatial location in the image. The system calculates the area of ​​the region, i.e., the number of pixels it contains, representing the spatial coverage of the region. The system extracts the minimum bounding rectangle (MBR) parameters of the region, including its width, height, and orientation angle in the image, representing the overall shape and arrangement of the region. The system extracts the contour of the boundary of each region, forming a continuous set of boundary points. The system calculates various region shape factors, such as roundness, aspect ratio, and boundary complexity (measured by calculating the total length of the polylines between boundary points and combining it with the frequency of boundary changes, such as the number of segmented angle changes or the rate of directional change per unit length of the boundary, thereby estimating the complexity of the region's edge). The system organizes the above information into regional spatial feature data, including region number, centroid location, area, minimum bounding rectangle parameters, shape indices, and boundary point set.

[0060] Step S22: Perform initial cross-channel regional matching based on regional spatial feature data to obtain regional matching data; In one embodiment, the system performs initial cross-channel region matching based on pre-generated regional spatial feature data. Starting with each target region in a reference channel (e.g., the ROS channel), the system searches for corresponding matching regions in other channels (e.g., iron ion channels and mitochondrial morphology channels). During the matching process, the system sets a series of criteria, such as requiring the distance between the centroid coordinates of the region to be matched and the centroid of the reference region to be less than a preset threshold (e.g., within 20 pixels); the area of ​​the matched region must be relatively consistent with the reference region, with the difference not exceeding a set ratio range (e.g., ±30%); and the system sets the aspect ratio difference between the matched region and the reference region to be small (e.g., controlled within 0.5). The system can match one or more candidate regions from other channels for each reference region and output them as a set of matching pairs. Each matching pair includes the reference region and its corresponding target channel region, resulting in region matching data.

[0061] Step S23: Calculate the matching offset for the region matching data to obtain the matching offset data; In one embodiment, the system calculates the offset between the centroid coordinates of the two regions to obtain matching offset data.

[0062] Step S24: Perform regional spatial offset correction on the image region data based on the matching offset data to obtain preliminary alignment data; In one embodiment, for each target channel, the system summarizes the centroid offsets of all matching pairs between the target channel and the reference channel. The system first performs validity screening based on the centroid distance, area difference, aspect ratio difference, and cross-union ratio (CUI) after translation of the matching regions, removing matching pairs with a centroid distance greater than 20 pixels, an area difference exceeding ±30%, an aspect ratio difference greater than 0.5, or an CUI lower than 0.10. The system calculates the median of the remaining valid offsets in the x and y directions to obtain an initial offset vector; it then calculates the deviation distance of each valid offset relative to the initial offset vector. If the deviation distance is greater than the median deviation distance plus three times the median absolute deviation, or greater than 15 pixels, it is identified as an outlier and excluded. The system recalculates the median of the remaining non-outlier offsets in the x and y directions to obtain the overall spatial offset vector of the target channel relative to the reference channel, and performs uniform translation correction on all image regions under the target channel according to this overall spatial offset vector to obtain preliminary alignment data.

[0063] Step S25: Perform fine-tuning of local spatial offset on the preliminary alignment data to obtain region alignment data.

[0064] In one embodiment, the system constructs a local registration window within the vicinity of each target region. This local registration window can be set to 30×30 pixels, or it can be obtained by extending the bounding rectangle of the target region outwards by 5 to 10 pixels. The system preferentially performs local translation fine-tuning based on the boundary point set between the reference channel region and the target channel region. Candidate offsets are generated in a step size of 1 pixel within a range of -3 to +3 pixels, and the candidate offset that minimizes the average nearest neighbor distance of the boundary and increases the region's intersection-union ratio (IU) or boundary overlap ratio is selected as the local correction offset. If there is slight deformation of the region boundary (e.g., the IU is not less than 0.20), the system performs local affine correction within a range of rotation angle not exceeding 5°, scale change ratio of 0.9 to 1.1, and translation amount not exceeding 5 pixels. If the affine correction does not improve the IU or reduce the boundary distance, the correction result is discarded. If the region boundary is blurred or there are insufficient contour points, the system instead performs a local offset search based on the consistency of the gray-level gradient direction or the mutual information value within the local window, and selects the candidate offset with the highest gradient consistency or the largest mutual information value as the local correction offset. The system updates the centroid, boundary, bounding rectangle, and mask coordinates of the target region based on the local correction offset. After fine adjustment, it calculates the centroid offset distance, the average nearest neighbor distance of the boundary, and the region intersection-union ratio. When the centroid offset distance is no greater than 3 pixels, the average nearest neighbor distance of the boundary is no greater than 3 pixels, and the region intersection-union ratio is no less than 0.30, the local fine adjustment is deemed effective, and the region alignment data is output.

[0065] Preferably, step S3 specifically includes: Step S31: Perform cross-channel correlation region filtering on the region alignment data to obtain cross-channel correlation region data; In one embodiment, the system constructs an extended neighborhood in the image space based on a region in a reference channel (such as the ROS channel). For example, it extends the outer boundary of the region by approximately 15 pixels and searches within this neighborhood for candidate regions from other channels. The system does not require complete overlap or strict geometric registration between regions. Instead, it determines whether any of the following conditions are met based on more relaxed spatial proximity and response coordination characteristics: First, there is a certain degree of boundary overlap between the reference region and regions in other channels, even if the overlap ratio is relatively small (e.g., approximately 10% to 20%); second, the region boundaries show a clear tendency to converge or approach each other in local locations, even if the overall structural morphology is somewhat different; third, the regions exhibit spatial clustering or edge contact characteristics in multiple channels, forming a weakly coupled cooperative response pattern. For example, within the extended neighborhood of the reference region, the system counts the number of regions from different channels and their spatial distribution. If there are two or more candidate regions from different channels within the extended neighborhood, and the centroids of all candidate regions fall within the extended neighborhood, the system further calculates the centroid distance between any two candidate regions, the distance from the centroid of each candidate region to the cluster center, and the nearest neighbor distance of each candidate region. When the centroid distance between any two candidate regions from different channels is no greater than 20 pixels, or no greater than the sum of their equivalent radii plus 5 pixels; and the maximum distance from the centroid of each candidate region to the cluster center is no greater than 15 pixels, or no greater than 50% of the equivalent diameter of the reference region; and there are no discrete regions where the distance from the centroid to the cluster center is greater than 25 pixels and the nearest distance to other regions is greater than 20 pixels, or the proportion of such discrete regions does not exceed 20%, the system determines that the spatial spacing between these regions is generally small and does not exhibit a significant discrete distribution. Therefore, it concludes that the response regions of multiple channels appear clustered around the same spatial location, exhibiting cross-channel clustering distribution characteristics. The system outputs all region pairs or region groups that meet the above conditions as cross-channel associated region data.

[0066] Step S32: Perform regional spatial consistency calculation based on the cross-channel associated regions to obtain regional spatial consistency data; In one embodiment, the system extracts the centroid position differences between associated regions and normalizes them. The system evaluates the spatial overlap of each region's boundaries by calculating the overlap ratio between regions to determine their spatial convergence. The system compares the differences in position and orientation angles of the bounding rectangles of each region. The system integrates the above multiple spatial feature indicators to form regional spatial consistency data.

[0067] Step S33: Determine the consistency of channel feature association based on the regional spatial consistency data to obtain channel feature association data; In one embodiment, the system filters cross-channel region groups that meet the condition of spatial consistency data, and extracts the average response intensity, relative response intensity, response uniformity index, and spatial gradient trend of the corresponding regions in the ROS channel, iron ion probe channel, and mitochondrial morphology channel, respectively. If the average response intensity of a candidate region in a certain channel is not less than the average background intensity plus twice the background standard deviation, or not less than 1.5 times the median intensity of the entire image of that channel, and the relative response intensity is not less than 20%, then the channel is determined to have an effective enhanced response. The system determines the response direction between channels. If at least two channels show effective enhancement, or their spatial gradient trend shows attenuation from the center outward in at least three consecutive radius layers, then the response change direction is determined to be consistent. The system also calculates the ratio of the pixel intensity standard deviation to the average response intensity in the region as a response uniformity index. When this index is not greater than 0.35, the response inside the region is determined to be stable; when this index is greater than 0.60 and only shows local isolated bright spots, local abnormal interference is determined to exist. The system calculates the average response intensity ratio among effective response channels. When this ratio is between 0.3 and 3.0, the channel response intensity ratio is considered to be within a reasonable range. When the ratio is greater than 5.0 and other channels do not meet the criteria for effective enhanced response, it is determined to be a single channel abnormally dominant. If at least two channels in a candidate region group have effective enhanced responses, and the response change direction is consistent, and no single channel abnormally dominates, the system determines that the region group has passed the channel feature association consistency judgment, and encapsulates the region group number, channel source, response intensity, gradient trend, uniformity index, and judgment result into channel feature association data.

[0068] Step S34: Construct multi-region association groups based on channel feature association data to obtain cross-channel association data.

[0069] In one embodiment, the system constructs a cross-channel region association graph based on channel feature association data. Candidate regions that pass the channel feature association consistency determination are used as graph nodes, and spatial proximity, boundary contact, and feature consistency relationships are used as association edges. If two candidate regions from different channels satisfy at least two of the following conditions: centroid distance is no greater than 20 pixels or no greater than the sum of their equivalent radii plus 5 pixels; region intersection-union ratio is no less than 0.10 or minimum boundary distance is no greater than 5 pixels; relative response intensity is no less than 20%; and spatial gradient trend remains consistent across at least three consecutive radius layers, then the system establishes a cross-channel association edge between them. The system determines the set of connected regions containing at least two different channel regions as candidate multi-region association groups and performs deduplication processing on duplicate regions within the same channel. If the maximum centroid distance of a candidate multi-region association group is greater than 30 pixels and there is no boundary contact relationship, or the ratio of the maximum region area to the minimum region area is greater than 3.0 and there is no nested inclusion relationship, then the system excludes it. For multi-region association groups that pass the verification, the system takes the union of the masks of each region in the group as the overall spatial coverage area, takes the geometric center of the boundary union as the central reference point, and encapsulates and outputs the association group number, participating channels, region number, spatial coverage range, boundary contact relationship and channel feature consistency results to obtain cross-channel association data.

[0070] Preferably, the confidence score is as follows: Step S41: Based on the cross-channel correlation data, perform regional alignment stability assessment, cross-channel support statistics, and channel change analysis to obtain regional alignment stability data, cross-channel support data, and channel consistency data, respectively. In one embodiment, the system performs region alignment stability assessment, cross-channel support statistics, and channel change analysis on each set of cross-channel associated data. For region alignment stability, the system calculates the centroid offset distance of the target channel region relative to the reference channel region, the average nearest neighbor distance, the region intersection-union ratio, and the centroid change before and after local fine-tuning. When the centroid offset distance is no greater than 3 pixels, the average nearest neighbor distance is no greater than 3 pixels, the region intersection-union ratio is no less than 0.30, and the centroid change is no greater than 5 pixels, the alignment stability is considered high. When the centroid offset distance is greater than 6 pixels or the region intersection-union ratio is less than 0.20, the alignment stability is considered low.

[0071] For cross-channel support, the system determines whether there is a valid response in each channel region. If the average response intensity of a region is not less than the average background intensity plus twice the background standard deviation, or not less than 1.5 times the median intensity of the entire image for that channel, and the relative response intensity is not less than 20%, the number of valid response pixels is not less than 30, and the proportion of valid response pixels is not less than 60%, then the channel is considered to have a valid response. When all three channels have valid responses, the support is considered high; when two channels have valid responses, the support is considered medium. For response distribution, the system uses the ratio of the standard deviation of pixel intensity within a region to the average response intensity as an indicator of response uniformity. When this ratio is not greater than 0.35, the response distribution is considered relatively uniform.

[0072] For channel variation analysis, the system determines the response direction based on the average response intensity, relative response intensity, and spatial gradient trend from the center of the region outwards for each channel region. If at least two channels show effective enhancement, and the relative response intensity is not less than 20%, and at least three consecutive radius layers show the same attenuation or diffusion direction, then the channel variation is considered consistent. If a channel's decrease exceeds 20% and its enhancement direction is opposite to that of other channels, then a directional conflict is considered. The system encapsulates the above results into region alignment stability data, cross-channel support data, and channel consistency data, respectively.

[0073] Step S42: Calculate the confidence level based on the regional alignment stability data, cross-channel support data, and channel consistency data to obtain the regional confidence level data.

[0074] In one embodiment, the system normalizes or transforms the three types of indicators mentioned above, dividing them into several confidence levels (e.g., high, medium, and low), with each level assigned a specific numerical value. The system combines and judges each indicator according to the following rules: When a region simultaneously meets the following three conditions—small spatial alignment residual, high channel support (e.g., significant response in all three channels), and consistent channel response trends (e.g., synchronous enhancement or diffusion)—the system marks the region as a high-confidence region, considering it to possess strong spatial reliability and consistency with biological responses, thus having high confidence. If a region performs well in two of the indicators (e.g., stable alignment and high support, but some differences in channel responses), it is judged as a medium-confidence region, possessing a certain degree of confidence but with some uncertainty. If a region performs well in only one indicator, or exhibits large spatial alignment deviations or inconsistent channel responses, the system marks it as a low-confidence region, using it cautiously as a weak candidate region. Through these rules, the system determines the confidence level of each region and outputs the corresponding region confidence data.

[0075] In one embodiment, another scoring method is given: alignment stability scoring: if the registration residual of the region in multiple channels is less than a set threshold (e.g., centroid offset < 5px, boundary Hausdorff distance < 10px), it is marked as "high" and assigned a value of 1.0; if there is a moderate residual (e.g., offset < 10px), it is marked as "medium" and assigned a value of 0.7; if the offset is large or the structural deformation is severe, it is marked as "low" and assigned a value of 0.4. Cross-channel support scoring: if the region responds in all three channels (average intensity is higher than 1.5 times the median value of each channel), it is marked as "high" and assigned a value of 1.0; if there is support in two channels, it is marked as "medium" and assigned a value of 0.7; if only one channel has a significant response, it is marked as "low" and assigned a value of 0.4. Channel Consistency Score: If the channel response trends are highly consistent (e.g., all three channels show enhancement or diffusion), it is marked as "High" and assigned a value of 1.0; if only two channels show consistency, or there is a slight trend difference, it is marked as "Medium" and assigned a value of 0.7; if the response directions are opposite and the changes are irregular, it is marked as "Low" and assigned a value of 0.4. The system averages the above three scores unweighted or slightly weighted (e.g., weight ratio 1:1:1 or 0.4:0.3:0.3) to obtain the confidence score for each region, with a confidence range of 0.0–1.0. Based on this score, the system classifies the confidence levels as follows: High confidence region: score ≥ 0.85; Medium confidence region: score in the range of 0.65–0.85; Low confidence region: score < 0.65.

[0076] Preferably, the pseudo-label output is as follows: Pseudo-label regions are filtered based on regional confidence data to obtain pseudo-label region data; In one embodiment, the system sets static or dynamic confidence screening criteria. For example, in static mode, the system presets a fixed confidence threshold (e.g., 0.85), and regions with confidence scores higher than this threshold are considered high-reliability regions. The system enables a dynamic threshold mechanism, adaptively selecting a subset of regions (e.g., the top 10% or Top-K regions) based on the confidence distribution of all regions in the entire image. The system automatically excludes all regions with confidence scores below the threshold to avoid potential label noise interfering with subsequent training or analysis. All regions meeting the screening criteria are output as pseudo-labeled region data, which are considered by the system as highly reliable positive examples without human intervention.

[0077] Perform region verification based on pseudo-label region data to obtain pseudo-label verification data; In one embodiment, the system performs structural stability verification, analyzing the morphological characteristics of the region across multiple channels. The system focuses on detecting obvious boundary incompleteness, local interruptions, or broken contours. For example, if the number of broken boundary segments is no less than three, or the effective closed contour length is less than 70% of the original contour length, or the number of skeleton branches in the region decreases by more than 40% compared to adjacent normal regions, it is considered a severe structural anomaly. If a region exhibits a severe structural anomaly in any channel, the system identifies it as an unstable region and removes it. The system performs response offset consistency verification to determine if the response center position of the region in each channel has shifted. If the response hotspot of a certain channel significantly deviates from the reference channel center beyond a set range (e.g., crossing regions or excessive offset), it indicates that the region is not a true co-localization response, and the system marks it as a spatially deviated region, excluding it from inclusion as a valid pseudo-label. The system also performs channel representation anomaly verification, checking whether a region has a strong signal only in one channel, while the signal is weak or even close to zero in other channels. Specifically, if the average response intensity of the corresponding region in the target channel is not higher than the mean of the background region plus one background standard deviation, or not higher than 1.1 times the midpoint value of that channel in the entire image, then the signal of that channel is determined to be close to zero. If it is confirmed that the region is dominated by a single channel and lacks cross-channel support, the system marks it as a biased region to avoid model learning bias caused by uneven structure of pseudo-labeled samples. All pseudo-labeled regions that pass the above verification process are officially recorded as pseudo-label verification data.

[0078] Based on the pseudo-label verification data, the pseudo-label category and labeling format are determined to obtain pseudo-label data; In one embodiment, the system analyzes the response patterns of pseudo-label regions in the ROS channel, iron ion probe channel, and mitochondrial morphology channel. For example, when a region exhibits a strong response in both the ROS and iron ion probe channels (the average pixel value / average intensity within the region exceeds 1.5 times the median value of the channel), accompanied by severe mitochondrial morphology damage (blurred boundary contours, fragmented regional structures, or reduced area and abnormal aspect ratio (less than or greater than a preset threshold) in the corresponding region within the mitochondrial morphology channel, while the signal intensity within the region is discontinuous or lower than the background average level), the system labels this region as a typical ferroptosis region, indicating that it conforms to the typical characteristics of ferroptosis in function and structure. If a region only shows significant enhancement in the ROS channel, with slight responses or blurred edges in other channels, it is judged as a suspected early region, indicating that it is in the early stage of ferroptosis. If a region shows similar response intensities and spatial distributions in all three channels, it is identified as a multi-channel co-occurrence region, used to capture interactive biological phenotypes. Based on the determination of the pseudo-label type, the system assigns a corresponding label to each region, supporting the combination and overlay of single or multiple types of labels. Meanwhile, based on specific application requirements and data usage scenarios, the system selects appropriate annotation formats, such as using masks to annotate pixel-level regions, or representing region boundary information in vector outlines, polygons, or other formats. The pseudo-label data output by the system includes structured content such as region location, label type, and annotation format.

[0079] Region annotations are generated based on pseudo-label data to obtain region annotation data.

[0080] In one embodiment, the system establishes an independent pseudo-label layer in the spatial dimension of the original image for overlaying and managing annotation information. For each pseudo-label region, the system draws it in the layer based on its boundary contour information and fills, colors, or labels it according to its category. For example, a region labeled "typical iron death zone" can be represented by a specific color and assigned a corresponding category label. When the system supports multi-level labeling or multi-channel annotation structures, it can also generate multi-channel composite images or output independent label images for each channel. The region annotation data generated by the system not only has a dual description of spatial location and category information, but can also be exported to commonly used image annotation standard formats, such as COCO, PascalVOC, and YOLO, facilitating integration into deep learning model training processes.

[0081] Preferably, this application also provides an automatic labeling system for ferroptosis cell regions, used to perform the automatic labeling method for ferroptosis cell regions as described above, the automatic labeling system for ferroptosis cell regions comprising: The region selection module is used to acquire multi-channel image data, which includes ROS channel data, iron ion probe channel data, and mitochondrial morphology channel data; and to select regions based on the multi-channel image data to obtain image region data. The spatial offset alignment module is used to perform spatial offset alignment based on image region data to obtain region alignment data. The cross-channel association module is used to perform cross-channel association on region-aligned data to obtain cross-channel associated data; The region labeling module is used to score confidence based on cross-channel correlation data to obtain region confidence data; and to output pseudo-labels based on the region confidence data to obtain region labeling data.

Claims

1. A method for automatically labeling regions of iron-dead cells, characterized in that, Includes the following steps: Step S1: Acquire multi-channel image data, which includes ROS channel data, iron ion probe channel data, and mitochondrial morphology channel data; select regions based on the multi-channel image data to obtain image region data; Step S2: Perform spatial offset alignment based on the image region data to obtain region alignment data; Step S3: Perform cross-channel correlation on the region-aligned data to obtain cross-channel correlated data; Step S4: Calculate the confidence score based on the cross-channel correlation data to obtain the regional confidence data; Pseudo-labels are output based on the regional confidence data to obtain regional labeled data.

2. The method according to claim 1, characterized in that, The specific steps for acquiring multi-channel image data are as follows: ROS channel data is acquired using the ROS imaging channel acquisition module; Data from the iron ion probe channel was acquired using the iron ion probe imaging module. Mitochondrial morphology channel data were acquired using a mitochondrial structure imaging module.

3. The method according to claim 1, characterized in that, The specific selection of the region is as follows: Preliminary regional screening is performed based on multi-channel image data to obtain preliminary regional screening data. Based on the initial regional screening data, cross-channel spatial neighborhood association is performed to obtain neighborhood association data; Weakly consistent region combination determination is performed on the neighboring related data to obtain image region data.

4. The method according to claim 3, characterized in that, The regional initial screening is specifically as follows: Cross-channel center point data is obtained by determining the center point across multiple channel image data. Intensity curves of channel spatial influence are constructed from cross-channel center point data to obtain intensity curve data; The diffusion radius of the channel is estimated based on the intensity curve data to obtain the diffusion radius data; Based on the diffusion radius data, the cross-channel diffusion radius relationship is determined, and the diffusion radius relationship data is obtained; Irrelevant diffusion regions are excluded from multi-channel image data based on diffusion radius relationship data to obtain preliminary regional screening data.

5. The method according to claim 3, characterized in that, The specific determination of weakly consistent region combinations is as follows: The minimum support across channels is determined for the neighborhood-related data to obtain primary regional unit data. Based on the spatial proximity consistency determination of the primary regional units, the secondary regional unit data is obtained; The consistency of regional scale relationships is determined by analyzing the secondary regional unit data to obtain the regional unit data. Image region data is obtained by hierarchically summarizing the primary region unit data, secondary region unit data, and region unit data.

6. The method according to claim 1, characterized in that, Step S2 is as follows: Regional spatial features are generated based on image region data to obtain regional spatial feature data; Based on the regional spatial feature data, initial cross-channel regional matching is performed to obtain regional matching data; The matching offset is calculated on the region matching data to obtain the matching offset data; Based on the matching offset data, the image region data is spatially offset corrected to obtain preliminary alignment data; Fine-tuning of the initial alignment data by local spatial offset is performed to obtain the region alignment data.

7. The method according to claim 1, characterized in that, Step S3 is as follows: Cross-channel associated region filtering is performed on the region-aligned data to obtain cross-channel associated region data; Regional spatial consistency data is obtained by performing regional spatial consistency calculations based on cross-channel associated regions; Based on the regional spatial consistency data, the consistency of channel feature association is determined, and the channel feature association data is obtained; Multi-region association groups are constructed based on channel feature association data to obtain cross-channel association data.

8. The method according to claim 1, characterized in that, The confidence score is as follows: Based on cross-channel correlation data, regional alignment stability assessment, cross-channel support statistics, and channel change analysis are performed to obtain regional alignment stability data, cross-channel support data, and channel consistency data, respectively. Confidence scores are calculated based on regional alignment stability data, cross-channel support data, and channel consistency data to obtain regional confidence scores.

9. The method according to claim 1, characterized in that, The pseudo-tag output is as follows: Pseudo-label regions are filtered based on regional confidence data to obtain pseudo-label region data; Perform region verification based on pseudo-label region data to obtain pseudo-label verification data; Based on the pseudo-label verification data, the pseudo-label category and labeling format are determined to obtain pseudo-label data; Region annotations are generated based on pseudo-label data to obtain region annotation data.

10. An automatic labeling system for ferroptosis cell regions, characterized in that, For performing the automatic ferroptosis cell region annotation method as described in claim 1, the automatic ferroptosis cell region annotation system comprises: The region selection module is used to acquire multi-channel image data, which includes ROS channel data, iron ion probe channel data, and mitochondrial morphology channel data; and to select regions based on the multi-channel image data to obtain image region data. The spatial offset alignment module is used to perform spatial offset alignment based on image region data to obtain region alignment data. The cross-channel association module is used to perform cross-channel association on region-aligned data to obtain cross-channel associated data; The region labeling module is used to score confidence based on cross-channel correlation data to obtain region confidence data; and to output pseudo-labels based on the region confidence data to obtain region labeling data.

Images