A medical image intelligent auxiliary decision-making method and system for perianal abscess and anal fistula

Abstract

The application discloses a perianal abscess and anal fistula medical image intelligent auxiliary decision method and system, and the intelligent auxiliary decision method comprises the following steps: collecting MRI sequence original images for pretreatment; obtaining clinical data for pretreatment; constructing a multi-task deep learning model, including: a feature extraction network layer, a region generation network, an output network layer, and a DAFT clinical information fusion module, extracting and fusing multi-parameter MRI image features and clinical data, and using the multi-task deep learning model to perform multi-task joint output of lesion automatic detection, internal orifice positioning, perianal abscess / anal fistula classification, and pathogenic bacteria classification.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and more specifically, to a medical imaging intelligent auxiliary decision-making method and system for perianal abscess and anal fistula. Background Technology

[0002] Perianal abscesses and anal fistulas are common diseases in anorectal surgery, characterized by complex anatomical structures, high postoperative recurrence rates, and the potential to damage anal function, seriously affecting patients' physical and mental health. Magnetic resonance imaging (MRI) is the first-line diagnostic method for perianal abscesses and anal fistulas, playing a crucial role in preoperative diagnosis, assessment, and treatment decision-making. However, current clinical interpretation of MRI findings in perianal diseases primarily relies on subjective visual assessment by physicians, resulting in low diagnostic efficiency, low homogeneity, and poor stability in identifying complex lesions. Furthermore, existing technologies lack products that provide intelligent image-assisted decision-making.

[0003] Given that key structures such as perianal abscesses, anal fistulas, and internal openings often occupy a small area in MRI images and may be accompanied by multi-branched fistulas, deep abscesses, and complex spatial relationships, relying solely on traditional manual image interpretation or single image task models is insufficient to simultaneously meet clinical needs such as automatic lesion detection, anal fistula classification, internal opening identification, and perianal abscess classification. Furthermore, the auxiliary diagnosis of perianal diseases depends not only on imaging morphological characteristics but also on patient clinical data and laboratory test results. Current technologies still do not fully utilize multi-parameter MRI image information and clinical data synergistically, limiting the accuracy and stability of auxiliary analysis results.

[0004] Prior art 1 (Publication No.: CN120565112A, Application No.: 2025106853557, Application Date: 2025.05.26) discloses a classification model, classification method, and apparatus based on multimodal multiscale fusion, specifically disclosing: acquiring a medical image to be detected and health data to be detected; extracting features from the medical image to be detected and the health data to be detected respectively to obtain visual features and health features; performing multiscale operations on the visual features and health features to obtain multiscale feature pairs; fusing features for each scale feature pair; and performing bidirectional multiscale fusion on the feature pairs after feature fusion to obtain multiscale fused features; and obtaining a classification result based on the multiscale fused features. However, the existing technology 1 has the following technical problems: it only achieves simple fusion of images and clinical data through one-dimensional vector cross attention, and the overall architecture is only suitable for general lung tumor single classification diagnosis scenarios; it can only achieve single tumor binary classification diagnosis, and cannot take into account lesion localization, segmentation, classification and pathogen tracing, and has poor recognition accuracy for perianal diseases with extremely high soft tissue precision requirements.

[0005] Prior art 2 (Publication No.: CN120257205A, Application No.: 2025103953093, Application Date: 2025.03.21) discloses a multimodal data dynamic fusion auxiliary prediction method and system for Crohn's fistula prediction, specifically disclosing: acquiring MRI images and clinical feature data, performing uncertainty estimation based on MRI images and uncertainty estimation based on clinical features, thereby realizing the uncertainty quantification of both MRI and clinical feature modalities, and then selecting the corresponding threshold strategy to measure credibility; adopting a multimodal dynamic fusion strategy, first calculating the credibility of the corresponding modality based on the corresponding threshold strategy, and then adaptively weighting the contributions of different modalities according to the credibility, performing multimodal dynamic fusion to improve the reliability and accuracy of the classification model in fistula prediction. ResNet50 or MobileNetV2 is used as the feature extraction network to extract features from MRI images, random forest is used to classify clinical features (patient age, BMI, hemoglobin, PLpR), and uncertainty estimation is combined to realize dynamic fusion decision of image modality and clinical feature modality. The existing technology 2 has the following problems: it only performs a single anal fistula classification task, without the ability to locate, segment, or quantify lesions, and can only output general classification results; it uses traditional CNN networks such as ResNet and MobileNet, which have poor ability to capture long-distance dependent features of small perianal fistulas, microabscesses, and superficial inflammation, and may miss a large number of subtle lesions; it can only achieve binary classification of the origin of anal fistula, and cannot take into account lesion location, segmentation, classification, and pathogen tracing.

[0006] Therefore, there is an urgent need to provide a technical solution that can improve the accuracy of imaging assessment of perianal diseases, improve the accuracy of auxiliary analysis results, and take into account lesion localization, segmentation, classification and pathogen tracing. Summary of the Invention

[0007] In view of this, the present invention provides a medical imaging intelligent auxiliary decision-making method and system for perianal abscess and anal fistula, which improves the accuracy of perianal disease imaging assessment and the accuracy of auxiliary analysis results, and enables the coordinated output of multiple tasks such as automatic lesion detection, anal fistula classification and internal opening identification, perianal abscess classification and pathogen classification.

[0008] By integrating image preprocessing, clinical data preprocessing, pyramid multi-scale feature extraction, and image and clinical feature fusion, the system achieves integrated output for tasks such as automatic lesion detection, anal fistula classification and internal opening identification, perianal abscess classification, and perianal abscess pathogen classification, providing technical support for standardized image assessment of perianal diseases.

[0009] In a first aspect, the present invention provides a medical imaging-assisted decision-making method for perianal abscess and anal fistula, comprising:

[0010] The original MRI sequence images were acquired, including T1-weighted imaging, T2-weighted imaging, and T2W1 fat-suppressed images. The original MRI sequence images were preprocessed to obtain preprocessed T1-weighted imaging, preprocessed T2-weighted imaging, and preprocessed T2W1 fat-suppressed images.

[0011] Acquire clinical data, which includes at least: demographic characteristics, medical history, relevant laboratory tests, anoscopy, pathogen culture and detection results, pathological results, and MRI diagnostic results. Preprocess the clinical data to obtain standardized clinical data.

[0012] A multi-task deep learning model is constructed, including: a feature extraction network layer, a region generation network, an output network layer, and a DAFT clinical information fusion module. This multi-task deep learning model is used for automatic lesion detection, internal opening localization, perianal abscess / anal fistula classification, and pathogen classification.

[0013] The feature extraction network layer includes a deep neural network, Swin Transformer-Tiny, and a feature pyramid network, FPN. It has multi-label, multi-channel input and uses Swin Transformer-Tiny and FPN to obtain feature maps from the three modalities: preprocessed T1-weighted imaging, preprocessed T2-weighted imaging, and preprocessed T2W1 fat-suppressed imaging. ;

[0014] The region generation network processes feature maps Generate candidate regions of interest to obtain candidate regions;

[0015] The output network layer includes a lesion segmentation branch, an internal opening localization branch, and a perianal abscess / anal fistula classification branch. An ROI Align module is set between the region generation network and the output network layer to extract a fixed-size region feature map from the candidate region.

[0016] The lesion segmentation branch is based on the output features of the ROI Align module, and uses the CBAM attention module to enhance the expression of spatial and channel information, and outputs a lesion segmentation mask;

[0017] The inner opening localization branch uses the Coord Attention module to enhance the coordinate attention of possible inner opening positions in the candidate region and output the inner opening heatmap or localization coordinates.

[0018] The classification branch uses the Squeeze-and-Excitation module to perform channel weighting on candidate regions or fusion features, and outputs the classification results of anal fistula and perianal abscess.

[0019] The pathogen classification task combines the features after fusion of clinical information to perform a three-class classification output, resulting in automatic lesion detection, internal opening localization, perianal abscess / anal fistula classification, and pathogen classification using the multi-task deep learning model.

[0020] The DAFT clinical information fusion module is used to dynamically adjust the feature fusion results of the standardized clinical data and input images, including:

[0021] Global average pooling is performed on the region feature map output by the ROI Align module to obtain a length of The channel-level global feature vector is then used to obtain the channel-level global feature vector; this vector is then combined with the clinical information vector. The features are concatenated to form a joint feature representation; the joint feature is then compressed through a first fully connected layer to... The dimension, after ReLU activation, is mapped to a length of [length missing] through a second fully connected layer. The vector, where the first Weiwei ,back Weiwei Among them, only for Apply activation function ,wrong Apply activation, and ultimately utilize and Channel-by-channel modulation is performed on the original region feature map to obtain a new feature map that incorporates clinical information. .

[0022] Optionally, preprocessing of the original MRI sequence images includes sequential format conversion, denoising, bias field correction, registration, resampling, intensity normalization, and region of interest cropping.

[0023] Optionally, feature maps are obtained from the three modalities of preprocessed T1-weighted imaging, preprocessed T2-weighted imaging, and preprocessed T2W1 fat-suppressed imaging using a deep neural network Swing Transformer-Tiny and a feature pyramid network FPN. ,include:

[0024] The three modalities—preprocessed T1-weighted imaging, preprocessed T2-weighted imaging, and preprocessed T2W1 fat-suppressed imaging—are respectively input into the feature extraction network layer, and the input set is... for:

[0025] ;

[0026] in, For preprocessed T1-weighted imaging, For preprocessed T2-weighted imaging, This is a pre-processed T2W1 fat-suppressed image;

[0027] After four stages of Swin Transformer-Tiny feature extraction, multi-scale feature maps of the three modalities at the four stages are obtained and output. Let the first stage be... The modal image in the first The feature maps output at each stage are:

[0028] ;

[0029] in, These correspond to preprocessed T1-weighted imaging, preprocessed T2-weighted imaging, and preprocessed T2W1 fat-suppressed imaging, respectively. This corresponds to stages 1 through 4; Indicates the first The mode in the th ... The Swing Transformer-Tiny feature extraction operation in each stage;

[0030] In each stage, a feature fusion module is set up between different modalities to perform cross-modal feature fusion of multimodal features at the same level. The outputs of the three modes in each stage are as follows: , , The feature map after feature fusion is represented as:

[0031] ;

[0032] in, This represents a feature concatenation operation along the channel dimension. Indicates the first Multimodal feature map after fusion of each stage;

[0033] After obtaining the feature fusion at each stage, a self-oriented feature pyramid is constructed using FNP. High-level feature maps are upsampled and then horizontally connected and added to adjacent low-level feature maps, forming a multi-scale feature representation that balances high-level semantic information and low-level spatial details. This includes setting the first... The features obtained after upsampling and lateral connection in each stage of the FPN path are: Then the first The final output features of each stage are represented as follows:

[0034] ;

[0035] in, This indicates element-wise addition. Indicates the first The feature maps output after each stage are fused by FPN are used as multi-scale inputs for subsequent task branches.

[0036] Optionally, the region generation network processes the feature map. Generate candidate regions of interest, resulting in candidate regions including:

[0037] Feature map output after FPN fusion A sliding window is used to pre-define several anchor boxes at each location. The classification score for each anchor box (whether it contains the target) and the regression offset from that anchor point to the true target box are predicted. The classification and regression are expressed as follows:

[0038] ;

[0039] ;

[0040] in, This indicates the probability that the candidate box is the target region. This represents the regression results for the candidate bounding box coordinates.

[0041] Optionally, a fixed-size region feature map is extracted from the candidate region, expressed as:

[0042] ;

[0043] in, This represents a fixed-size feature map extracted from the candidate region. This indicates the coordinate information of the candidate region.

[0044] Optionally, an uncertainty-weighted strategy is used to jointly optimize the multi-task loss, where the lesion segmentation loss is denoted as . The internal positioning loss is Fractal loss is The total loss can then be expressed as:

[0045] ;

[0046] in, For the task The learnable uncertainty parameter.

[0047] Optionally, the DAFT clinical information fusion module is used to dynamically adjust the feature fusion results of the standardized clinical data and input images, including:

[0048] Suppose that for the i-th segmentation instance in the dataset, its clinical information vector Represented as:

[0049] ;

[0050] Its ROI-level image feature map Represented as:

[0051] ;

[0052] in, Representing the channel index, the DAFT feature vector fusion module uses an auxiliary network. Scaling factor is generated based on image features and clinical information. and offset Then, a channel-by-channel affine transformation is performed on the original feature map:

[0053] ;

[0054] ;

[0055] ;

[0056] in, and This is a mapping function modeled by an auxiliary neural network.

[0057] Secondly, this invention provides a medical imaging intelligent auxiliary decision-making system for perianal abscess and anal fistula, applied to the aforementioned auxiliary decision-making method, comprising: an image preprocessing module, a clinical data preprocessing module, a pyramid feature fusion model for multimodal images, a DAFT clinical information fusion module, a region generation network, and an ROI Align module, wherein...

[0058] The image preprocessing module is used to preprocess the acquired MRI sequence raw images to obtain preprocessed T1-weighted images, preprocessed T2-weighted images, and preprocessed T2W1 fat-suppressed images.

[0059] The clinical data preprocessing module is used to preprocess the clinical data to obtain standardized clinical data.

[0060] The pyramid feature fusion model for multimodal images is used to extract multi-scale image features;

[0061] The DAFT clinical information fusion module is used to fuse clinical information with image features from multiple MRI sequences.

[0062] After fusing clinical information and imaging features from multiple MRI sequences, the data is processed by the region generation network and the ROIAlign module to output branches for automatic lesion detection, internal opening localization, perianal abscess / anal fistula classification, and pathogen classification.

[0063] Thirdly, the present invention also provides an MRI device including the above-described auxiliary decision-making system and performing the above-described auxiliary decision-making method.

[0064] The multi-task deep learning model in this invention uses the Swin Transformer-Tiny as its backbone, combined with a Feature Pyramid Network (FPN), to achieve unified modeling of lesion structures at different scales. This structure balances long-range dependency modeling capabilities with multi-scale feature fusion capabilities, enabling simultaneous identification of key targets such as large abscess structures and minute internal openings. In terms of multi-task collaboration, the model achieves multiple key tasks based on shared backbone features, including lesion segmentation, disease classification, internal opening localization, and pathogen classification. To address the differences in information focus across tasks, a task-oriented attention mechanism is constructed, and an uncertainty-weighted strategy is used to dynamically balance multi-task losses, improving training stability and overall performance. The model output focuses on key structures and diagnostic elements in imaging, and is comprehensively evaluated based on indicators such as segmentation accuracy, classification accuracy, and localization error. Interpretability analysis is incorporated to ensure the clinical credibility of the results.

[0065] Compared with existing technologies, the intelligent assisted decision-making method and system for medical imaging of perianal abscess and anal fistula provided by this invention achieves at least the following beneficial effects:

[0066] This invention constructs a multi-task deep learning model for perianal abscesses and anal fistulas, capable of extracting and fusing multi-parameter MRI image features and clinical data. It achieves collaborative output across multiple tasks, including automatic lesion detection, anal fistula classification and internal opening identification, and perianal abscess classification and pathogen classification. Compared to existing methods that primarily rely on manual image interpretation or single-task analysis, this invention improves the efficiency, accuracy, and homogenization of perianal disease image assessment, and enhances the ability to identify complex lesions and small target structures.

[0067] Compared with existing technologies, this invention does not employ a dual-module coarse and fine segmentation technique based on two-dimensional slices along the main direction, nor does it use anorectal disease diagnostic models primarily based on U-net, 2.5D models, or 3D CNNs. Instead, it constructs a multimodal feature extraction framework combining a Swing Transformer-Tiny network with a feature pyramid network based on three MRI sequences: T1WI, T2WI, and T2WI_FS. Candidate region features are extracted through a region generation network and ROI Align. Combined with a DAFT clinical information fusion module, dynamic affine fusion of clinical information and imaging features is achieved. Ultimately, within a unified framework, multiple tasks are jointly output, including automatic lesion detection, internal opening localization, anal fistula classification, perianal abscess classification, and pathogen classification. This improves the accuracy, consistency, and clinical applicability of preoperative imaging assessment for perianal diseases.

[0068] Of course, any product implementing this invention does not necessarily need to achieve all of the technical effects described above at the same time.

[0069] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0070] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0071] Figure 1 This is a flowchart of the intelligent assisted decision-making method for medical imaging of perianal abscess and anal fistula provided by the present invention;

[0072] Figure 2 This is a general structural framework diagram of the present invention;

[0073] Figure 3 This is a structural diagram of the multi-task deep learning model of the present invention;

[0074] Figure 4 This is a diagram of the multimodal feature pyramid fusion structure of the present invention;

[0075] Figure 5 This is a structural diagram of the dynamic affine feature fusion module of the present invention. Detailed Implementation

[0076] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0077] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0078] Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and apparatus should be considered part of the specification.

[0079] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0080] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0081] For ease of understanding, the data acquisition method in this invention will be explained in detail below.

[0082] (1) Research Sample and Data Collection

[0083] This invention employs a retrospective case collection method. The study sample comprises clinical, imaging, and laboratory data from previously hospitalized patients, without involving public recruitment of volunteers. Included cases are all patient data from hospital medical record systems, image archiving and transmission systems, laboratory information systems, and pathology systems, representing patients who have completed their treatment processes. The collection, organization, and analysis of relevant data are conducted in accordance with hospital ethical approval and data anonymization requirements.

[0084] The inclusion criteria for the study sample were: age ≥18 years; meeting the diagnostic criteria in the "Chinese Expert Consensus on Clinical Diagnosis and Treatment of Perianal Abscess 2018"; having no contraindications for MRI examination; and having completed MRI examination before surgery.

[0085] Baseline data collected upon admission included:

[0086] 1) Demographic characteristics: age, sex, height, weight, body mass index;

[0087] 2) Medical history: onset time, main symptoms, history of perianal abscess or anal fistula, history of perianal surgery, history of diabetes, history of inflammatory bowel disease, smoking history, alcohol consumption history and other relevant underlying diseases;

[0088] 3) Relevant laboratory tests: routine blood tests (such as white blood cell count and neutrophil count), inflammatory markers (such as C-reactive protein and procalcitonin), biochemical indicators (such as blood glucose) and other routine clinical tests;

[0089] 4) Anoscopy data: preliminary location of the internal opening, condition of the perianal mucosa, purulent discharge, and visible fistula opening.

[0090] Collection of abscess and anal fistula specimens: Assume a suitable patient position, such as lateral decubitus. Disinfect the perianal area with povidone-iodine, drape the area, and strictly adhere to aseptic techniques. After surgical incision of the abscess lesion, use a disposable syringe to aspirate 2–5 ml of pus and send it to the microbiology laboratory of the clinical laboratory; excised anal fistula tissue is sent to the pathology department.

[0091] Pathogen culture and mass spectrometry detection: Collected pus samples were inoculated onto blood agar plates and MacConkey agar plates, incubated at 35°C for 18–24 h, and identified using the VITEK-2 Compact fully automated microbial identification system from bioMérieux, France. The cultured samples were then placed in a mass spectrometer, and the MALDI-TOF method was used to obtain microbial protein fingerprints. These fingerprints were then compared with a database to quickly determine the microbial identification results.

[0092] Histopathological processing of anal fistula tissue: The excised tissue was embedded in paraffin to form a paraffin block, which was then cut into 4-6 μm thin sections on a microtome. The sections were then spread out on glass slides and stained with hematoxylin and eosin (HE). Finally, pathological diagnosis was performed under a microscope.

[0093] (2) Standardized acquisition of MRI sequence images

[0094] The MRI examination employed a small field of view (FOV). The scanning instrument was a Canon Vantage Titan 3.0T superconducting MRI scanner with an abdominal coil. No special bowel preparation or antispasmodics was required before the examination. The patient was positioned prone, and a uniform scanning baseline was used: the transverse plane was perpendicular to the long axis of the anal canal, while the coronal and sagittal planes were parallel to the long axis of the anal canal. The examination protocol was standardized, and the scanning sequences included transverse T1WI / TSE, T2WI / TSE, and T2WI fat-suppressed sequences. Finally, all acquired MRI images were uploaded to the image workstation terminal (Winning Health TView, version 6.1.0) of the Picture Archiving and Communication Systems (PACS) system. The specific scanning protocol is shown in Table 1.

[0095] Table 1. MRI sequence parameters of the anal canal

[0096]

[0097] It should be noted that in Table 1: T1WI represents T1-weighted imaging; T2WI represents T2-weighted imaging; FS represents fat suppression; TR represents repetition time; TE represents echo time; FOV represents field of view; FSE represents fast spin echo; SPAIR represents spectral attenuated inversion recovery; and SPEEDER represents parallel acquisition parameters.

[0098] (3) Classification and grading assessment of anal fistula and perianal abscess, and classification of pathogens causing perianal abscess.

[0099] Anal fistulas and perianal abscesses are usually demarcated by a 1cm diameter boundary. The internal opening of anal fistula is located using the lithotomy position clock method, which divides the anus into 12 equal parts, corresponding to the time on a clock, and recorded as "1 o'clock" to "12 o'clock".

[0100] Anal fistula classification methods include the SPTF classification and the St. James University Hospital classification. Anal fistulas are classified into simple and complex anal fistulas based on complexity (SPTF classification, developed by the American College of Colon and Rectal Surgeons), as follows: Simple anal fistulas generally have a clear and simple structure, consisting of a single fistula tract with only one internal and one external opening; complex anal fistulas have a more complex structure, with multiple fistula tracts and multiple internal or external openings. Another MRI-based classification method is more suitable for assessing the degree and extent of the lesion, namely the St. James University Hospital classification, with the following specific grading criteria: Grade 0, normal; Grade 1, simple intersphincteric anal fistula; Grade 2, intersphincteric anal fistula with intersphincteric abscess or secondary sinus tract; Grade 3, transsphincteric; Grade 4, transsphincteric with ischioanal fossa or ischiorectal fossa abscess or secondary sinus tract; Grade 5: supralevator ani or translevator ani type.

[0101] Classification of perianal abscesses: Based on MRI images, the abscesses are classified according to the levator ani muscle at the upper edge of the anal canal (abscesses above the levator ani muscle are deep abscesses, and those below the levator ani muscle are superficial abscesses).

[0102] Classification of pathogens causing perianal abscesses: They are divided into three categories, namely Escherichia coli, Klebsiella pneumoniae, and others.

[0103] (4) Dataset Construction

[0104] Study subjects were collected according to the same inclusion criteria, and the preoperative MRI examination protocol was uniform and standardized. During the operation, abscess specimens were collected and sent to the microbiology laboratory of the laboratory department for pathogen culture and mass spectrometry detection on time, and the excised tissue was sent to the pathology department for pathological sectioning.

[0105] Establish a data table and enter baseline data upon admission (including demographic characteristics, medical history, relevant laboratory tests, anoscopy, etc.), pathogen culture and detection results, pathological results, and MRI diagnostic results (including classification of perianal abscess and anal fistula, location of internal opening, etc.).

[0106] The data sources for this invention are divided into three parts: 1) Training set: Data collected from a tertiary hospital (hereinafter referred to as "our hospital") between August 2019 and April 2024, including 520 cases of perianal abscess (including secondary anal fistula) and 400 cases of simple anal fistula; 2) Internal validation set: Data collected from our hospital between May 2024 and February 2025, including 150 cases of perianal abscess (including secondary anal fistula) and 100 cases of simple anal fistula; 3) External validation set: Data from two other tertiary hospitals, including 300 cases of perianal abscess and 300 cases of anal fistula. Pathogen species data were all from our hospital, totaling 220 cases.

[0107] Example 1

[0108] In examples of the present invention, such as Figure 1 and Figure 2 As shown, this invention provides an intelligent auxiliary decision-making method for medical images of perianal abscesses and anal fistulas. It utilizes a Swin Transformer-Tiny multi-task deep learning model (Feature Pyramid Module based Multi-task Learning Model, FP-MTL) for joint modeling of multi-parameter MRI images and clinical data. FP-MTL employs DAFT technology to fuse clinical information with image features output from the Swin Transformer-Tiny feature pyramid network, combining multi-channel MRI data to achieve multi-classification tasks related to perianal abscesses and anal fistulas.

[0109] Based on multi-parameter MRI image preprocessing and clinical data preprocessing, this invention achieves functions such as automatic lesion detection, anal fistula classification and internal opening identification, perianal abscess classification and perianal abscess pathogen classification through multi-scale feature extraction and image-clinical information fusion, thereby improving the accuracy, consistency and application efficiency of preoperative image assessment.

[0110] S101, acquire raw MRI sequence images, including T1-weighted imaging, T2-weighted imaging and T2W1 fat-suppressed images. Preprocess the raw MRI sequence images to obtain preprocessed T1-weighted imaging, preprocessed T2-weighted imaging and preprocessed T2W1 fat-suppressed images.

[0111] In this embodiment, MRI sequence images are acquired and preprocessed in a standardized manner. The MRI examination employs a small field of view (FOV) and an abdominal coil. No special bowel preparation or antispasmodic medication is required before the examination. The patient is positioned prone, and a unified scanning baseline is used: the transverse plane is perpendicular to the long axis of the anal canal, while the coronal and sagittal planes are parallel to the long axis of the anal canal. All examination protocols are standardized, and the scanning sequences include T1WI, T2WI, and T2WI_FS sequences. This patent is the first to use multi-parameter MRI sequences simultaneously to construct an intelligent decision-making system for perianal abscess and anal fistula. The three types of MRI data are input via three channels, and multiple labels are assigned to each input sample based on the needs of multi-classification tasks. To ensure consistent spatial correspondence between different sequences and improve the stability of subsequent feature extraction, the acquired MRI images undergo sequential format conversion, denoising, bias field correction, registration, resampling, intensity standardization, and region of interest cropping.

[0112] Specifically, the DICOM format images exported from the PACS system are first converted to NIfTI format and archived according to patient number and sequence type. Then, each sequence image is denoised, preferably using nonlocal mean filtering or Gaussian smoothing filtering to suppress random noise and preserve lesion edge information. Afterward, bias field correction is performed on each sequence image to eliminate low-frequency grayscale drift in MRI images caused by magnetic field inhomogeneities.

[0113] Furthermore, using the T2WI sequence as a reference sequence, the T1WI and T2WI_FS sequences were registered. The registration process preferably employed a rigid registration method based on mutual information, first optimizing translation and rotation parameters to achieve spatial alignment of different sequences in the anal canal and perianal region. When significant scale differences existed, affine registration was further used for fine correction. After registration, each sequence image was resampled to ensure uniform voxel spacing and matrix size. Specifically, trilinear interpolation was preferably used for image data resampling, and nearest-neighbor interpolation was preferably used for the annotation mask to avoid distortion of boundary information.

[0114] After spatial alignment, intensity normalization is performed on each sequence of images. Preferably, the background area is removed first, leaving only the effective signal area of ​​the human body; then the gray values ​​of each sequence are truncated, and abnormal gray values ​​below a preset lower limit and above a preset upper limit are removed; subsequently, the Z-score normalization method is used to normalize the gray values ​​of each sequence of images to a distribution with a mean of 0 and a standard deviation of 1, or the min-max normalization method is used to map the gray values ​​to the [0,1] interval, so as to reduce the gray value differences between different patients and different scanning batches.

[0115] To enhance the model's focus on localized perianal lesions, regions of interest (ROIs) were cropped after image standardization. Specifically, the cropping range was determined on the reference T2WI sequence, centered on the anal canal and surrounding tissues. The same spatial coordinates were then simultaneously applied to the T1WI and T2WI_FS sequences, resulting in multi-sequence local image patches containing perianal abscesses, fistulas, and related anatomical structures such as internal openings. These patches served as subsequent network input. During network training, further data augmentation of the cropped images was performed, including random rotation, translation, scaling, flipping, and contrast perturbation, to improve the model's robustness and generalization ability.

[0116] After the above preprocessing, the three MRI sequence images are input into the multi-task deep learning model in a multi-channel manner, and combined with the corresponding multi-label annotation information, to complete tasks such as automatic lesion detection, anal fistula classification and internal opening identification, perianal abscess classification and perianal abscess pathogen classification.

[0117] S102. Obtain clinical data, which shall include at least: demographic characteristics, medical history, relevant laboratory tests, anoscopy, pathogen culture and detection results, pathological results, and MRI diagnostic results. Preprocess the clinical data to obtain standardized clinical data.

[0118] In this embodiment, the preprocessing of clinical data includes data cleaning, field standardization, text structuring, missing value handling, outlier handling, numerical standardization, categorical variable encoding, and multi-source data association matching.

[0119] Specifically, clinical data records are first established based on the patient's unique identifier. The demographic characteristics, medical history, relevant laboratory test results, anoscopy results, pathogen culture and detection results, pathological results, and MRI diagnostic results of the same patient are then linked and organized in chronological order according to the examination time.

[0120] Secondly, the clinical data is cleaned to remove duplicate records, invalid records, and records that are irrelevant to the decision-making process for perianal abscess or anal fistula; consistency checks are performed on fields such as patient number, gender, age, examination date, test item name, and diagnosis conclusion, and data with obvious entry errors or formatting errors is corrected.

[0121] Furthermore, clinical data from different sources are standardized by mapping demographic characteristics, medical history, relevant laboratory tests, anoscopy, pathogen culture and detection results, pathological results, and MRI diagnostic results to preset field templates. These preset field templates include at least the following: age, sex, body mass index, onset time, history of perianal abscess or fistula, history of perianal surgery, history of diabetes, history of inflammatory bowel disease, smoking history, alcohol consumption history, white blood cell count, neutrophil count, C-reactive protein, procalcitonin, blood glucose, location of the internal opening under anoscopy, purulent discharge, fistula opening, pathogen type, pathological diagnosis results, MRI diagnostic classification, and location of the internal opening.

[0122] For relevant laboratory test results, the names of test items, units of measurement, numerical formats and reference ranges from different data sources are standardized, and the same test index is converted into a unified expression form; for the text content of anoscopy results, pathogen culture and detection results, pathological results and MRI diagnostic results, the lesion site, internal opening location, fistula opening status, pathogen name, pathological diagnosis conclusion, anal fistula type, perianal abscess type and MRI abnormal signs are extracted and converted into structured fields.

[0123] For numerical clinical data, outliers exceeding the preset physiological or statistical range are identified and corrected, truncated, removed, or retained according to preset rules. For missing data, processing is performed based on field type and degree of missingness. Specifically, when key diagnostic fields are missing, the sample is removed or marked. Non-key continuous fields are filled with the mean, median, or preset value, and non-key categorical fields are filled with the mode, unknown category, or missing marker.

[0124] For continuous data, Z-score standardization was used for numerical standardization; for categorical data, one-hot coding, label coding, or binary coding methods were used for coding. Among them, the presence or absence of past medical history, diabetes history, inflammatory bowel disease history, smoking history, and drinking history in the medical history data were converted into binary variables, and the classification of anal fistula, perianal abscess, pathogen type, and MRI diagnosis results were converted into categorical variables.

[0125] Finally, based on the patient's unique identifier and examination time, the preprocessed clinical data is matched with the corresponding MRI sequence images, and the preprocessed clinical fields are combined to form a clinical information vector, which serves as the input for the subsequent DAFT clinical information fusion module, thereby obtaining standardized clinical data.

[0126] S103, Construct a multi-task deep learning model, including: a feature extraction network layer, a region generation network, an output network layer, and a DAFT clinical information fusion module. Use the multi-task deep learning model to perform automatic lesion detection, internal opening localization, perianal abscess / anal fistula classification, and pathogen classification.

[0127] In this embodiment of the invention, the first part is a feature extraction network layer, i.e., as shown below. Figure 4 The multimodal image pyramid feature fusion model 30 shown employs the Swin Transformer-Tiny as the backbone network and combines it with Feature Pyramid Networks (FPN) to construct a multimodal, multi-scale feature extraction framework. This structure models long-distance spatial dependencies through a hierarchical window attention mechanism while preserving local structural details, effectively identifying complex anatomical relationships and small lesion structures in the perianal region.

[0128] Assuming the three input MRI sequences are T1-weighted imaging, T2-weighted imaging, and fat-suppressed T2-weighted imaging, the input set can be represented as follows:

[0129] ;

[0130] in, This is the preprocessed T1-weighted imaging, i.e., T1WI image. This is the preprocessed T2-weighted imaging, i.e., T2WI image. This is the preprocessed T2W1 fat-suppressed image, i.e., the T2WI_FS image.

[0131] Each input image has a uniform spatial size after preprocessing; for example, in one embodiment, it can be cropped and resampled. 3D image blocks.

[0132] The pyramid feature fusion model 30 for multimodal images mainly consists of two parts: the first part is a bottom-up feature extraction path based on the SwinTransformer-Tiny feature extraction module 31, and the second part is a top-down multi-scale fusion path based on the feature pyramid network 32 (FPN). In the bottom-up path, each modal image goes through four stages of the Swin Transformer-Tiny feature extraction module 31, outputting feature maps at different scales. Let the first... The modal image in the first The feature maps output at each stage are:

[0133] ;

[0134] in, These correspond to T1WI, T2WI, and T2WI_FS, respectively. ,correspond Figure 4 Phases 1 to 4 of the process; Indicates the first The mode in the th ... The Swin Transformer-Tiny feature extraction operation is performed in four stages. This yields multi-scale feature maps of three modalities across four stages.

[0135] like Figure 4 As shown, in each stage, a feature fusion module is set up between different modalities to perform cross-modal fusion of multimodal features at the same level. Let the first... The outputs of the three modes of the phase are respectively , and Then the feature map after fusion at this stage is represented as:

[0136] ;

[0137] in, This represents a feature concatenation operation along the channel dimension. Indicates the first The multimodal feature map is fused from each stage. The purpose of this step is to integrate the complementary information provided by different MRI sequences. For example, T2WI is more sensitive to displaying fluid lesions, T2WI_FS displays inflammation and lesion boundaries more clearly, and T1WI helps to provide anatomical background information, thereby enhancing the comprehensive characterization of lesion morphology and internal structure.

[0138] After obtaining the fused features at each stage, a top-down feature pyramid is constructed using FPN. Specifically, high-level feature maps are upsampled, then horizontally connected to and added to adjacent low-level feature maps to form a multi-scale feature representation that takes into account both high-level semantic information and low-level spatial details. Let the... The features obtained after upsampling and lateral connection in the stage FPN path are: Then the first The final output feature of the stage can be represented as:

[0139] ;

[0140] in, This indicates element-wise addition. Indicates the first The feature maps output after fusion by FPN at each stage. The FPN ultimately outputs fused feature layers of different scales, which in this embodiment can be represented as follows: , and .in, Features corresponding to higher resolution and richer local details Features that correspond to stronger semantic information. Figure 4 The right side shows , , This serves as the multi-scale input for subsequent task branches.

[0141] It should be noted that, Figure 3 The "feature map" mentioned above refers to the feature map set obtained after multimodal feature extraction and FPN fusion. Or its corresponding scale feature layer. In other words, Figure 3 The inputs sent to the RPN and ROI Align modules originate from... Figure 4 The output is the fused multi-scale features.

[0142] The second part is layer 50 of the Region Proposal Network (RPN). The RPN's role is to generate candidate regions of interest on the feature map to address the problems of large scale variations and irregular shapes in lesion targets. The RPN is not equivalent to an "anchor mechanism," but rather an end-to-end trainable candidate region generation network; the anchor mechanism is merely an implementation within the RPN used to generate candidate boxes of different scales and aspect ratios. Specifically, the RPN generates candidate boxes on the feature map... A sliding window is used to pre-define several anchor boxes at each location. The classification score for each anchor box (whether it contains the target) and the regression offset from that anchor box to the true target box are then predicted. The classification and regression can be represented as follows:

[0143] ;

[0144] ;

[0145] in, This indicates the probability that the candidate box is the target region. This represents the regression results for the candidate bounding box coordinates.

[0146] In one specific embodiment, assume the size of the input fused feature map is... If three scales and three aspect ratios are set at each spatial location on this feature map, resulting in a total of nine anchor points, then the RPN output includes two parts: one is the classification probability for each anchor point, such as... Secondly, the regression parameters corresponding to each anchor point. After filtering, the highest-scoring candidate boxes are retained; for example, a candidate region can be obtained. This is used for subsequent fine-grained classification and segmentation.

[0147] The third part is the output network layer. This part corresponds to... Figure 3 The multi-task output region, indicated by the blue dashed box, includes branches for lesion segmentation, internal opening localization, and perianal abscess / fistula classification. A ROI Align module 60 is placed between the RPN and the output network layer to extract fixed-size region feature maps from candidate regions. Its expression is:

[0148] ;

[0149] in, This represents a fixed-size feature map extracted from the candidate region. This represents the coordinate information of the candidate region. ROI Align is an operation used in deep learning to process Region of Interest (ROI). The improvement of the ROI Align module compared to traditional ROI Pooling lies in: eliminating coordinate discretization and quantization rounding operations, and using bilinear interpolation to sample features within the candidate region, thereby reducing boundary offset errors and improving the localization accuracy of small target structures, especially internal openings and small fistula branches.

[0150] In this invention, the lesion segmentation branch, based on the output features of the ROI Align module 60, employs the CBAM attention module 70 to enhance spatial and channel information representation and outputs a lesion segmentation mask. Specifically, the segmentation branch predicts the probability of each pixel location belonging to a lesion or background, and finally obtains a binary segmentation result after thresholding, achieving automatic segmentation of perianal abscesses, fistulas, and related lesion areas. The internal opening localization branch uses the CoordAttention module 80 to enhance the coordinate attention of possible internal opening locations in the candidate region and outputs an internal opening heatmap or localization coordinates; the classification branch uses the Squeeze-and-Excitation module 90 to perform channel weighting on the candidate region or fused features, outputting the classification results of anal fistulas and perianal abscesses. For the pathogen classification task, the features after fusion of clinical information are further combined to perform a three-class classification output, obtaining the probabilities of Escherichia coli, Klebsiella pneumoniae, and other classes.

[0151] To balance the differences in learning difficulty and gradient scale among different tasks, this invention employs an uncertainty weighting strategy to jointly optimize the multi-task loss. Let the lesion segmentation loss be... The internal positioning loss is Fractal loss is The total loss can then be expressed as:

[0152] ;

[0153] in, Task The method employs learnable uncertainty parameters. It can automatically adjust weights based on the noise level and optimization difficulty of different tasks during training, preventing any single task from dominating the training process and improving overall convergence stability and generalization ability.

[0154] Furthermore, the present invention also includes a DAFT clinical information fusion module 40, where DAFT stands for Dynamic Affine Feature Map Transformation, referring to... Figures 2 to 5 This is the fourth part of the network. This module corresponds to... Figure 5 The structure shown can dynamically adjust the feature fusion effect based on images and clinical data. It is important to clarify that the "image" input to the DAFT clinical information fusion module 40 is not the original MRI image, but rather a processed image. Figure 4 Multimodal feature extraction, Figure 3 The region feature map obtained after RPN and module processing Therefore, the object of global average pooling in the DAFT clinical information fusion module 40 is the feature map output by the ROI Align module 60, not the original image, nor the initial features that have not yet entered the candidate region processing. Figure 3 The connecting line to the right of the DAFT clinical information fusion module 40 indicates that DAFT operates on ROI-level regional features. DAFT technology is a computing engine for DATA + AI multimodal data processing and analysis scenarios. It supports both standalone and distributed operating modes, its kernel is written in Rust, and it provides both SQL and Python DataFrame interaction methods.

[0155] Suppose that for the i-th segmentation instance in the dataset, its clinical information vector is represented as:

[0156] ;

[0157] Its ROI-level image feature map is represented as follows:

[0158] ;

[0159] in, This indicates the channel index. The DAFT clinical information fusion module 40 uses an auxiliary network. Scaling factor is generated based on image features and clinical information. and offset Then, a channel-by-channel affine transformation is performed on the original feature map:

[0160] ;

[0161] ;

[0162] ;

[0163] in, and This is a mapping function modeled by an auxiliary neural network.

[0164] In practice, the feature map output by ROI Align is first subjected to global average pooling to obtain a length of... The channel-level global feature vector is then used to obtain the channel-level global feature vector; this vector is then combined with the clinical information vector. The features are concatenated to form a joint feature representation; the joint feature is then compressed through a first fully connected layer to... The dimension, after ReLU activation, is mapped to a length of [length missing] through a second fully connected layer. The vector, where the first Weiwei ,back Weiwei Among them, only for Apply activation function , and not Apply activation. Ultimately utilize. and Channel-by-channel modulation is performed on the original region feature map to obtain a new feature map that incorporates clinical information. .

[0165] In one specific embodiment, let the size of the output feature map of a certain ROI Align be... If the global average pooling is applied, the resulting vector will be 256 in length; if the patient's clinical information vector has a dimension of 9, the concatenated vector will have a dimension of 265. Let the compression factor be... The first fully connected layer outputs a vector of length 53, and after passing through the second fully connected layer, it outputs a vector of length 512, where the first 256 dimensions are the scaling parameters. The last 256 dimensions are offset parameters. For example, the parameters for some channels can be:

[0166] , , ,

[0167] , , ,

[0168] The corresponding channel transformation result is:

[0169] ;

[0170] ;

[0171] ;

[0172] This enables dynamic modulation of clinical information on the expression of imaging features.

[0173] Through the aforementioned structural design, this invention achieves efficient and accurate detection and segmentation of perianal abscesses and anal fistulas, providing a powerful auxiliary tool for clinical diagnosis and treatment. Furthermore, the system of this invention possesses high scalability and adaptability, and can be applied to other types of medical image analysis tasks.

[0174] Example 2

[0175] This embodiment provides an intelligent decision-making system for perianal abscess and anal fistula, combined with Figures 2 to 5 It includes: an image preprocessing module 10, a clinical data preprocessing module 20, a multimodal image pyramid feature fusion model 30, a DAFT clinical information fusion module 40, a region generation network 50, and an ROI Align module 60, among which,

[0176] The image preprocessing module 10 is used to preprocess the acquired MRI sequence raw images to obtain preprocessed T1-weighted images, preprocessed T2-weighted images, and preprocessed T2W1 fat-suppressed images.

[0177] The clinical data preprocessing module 20 is used to preprocess the clinical data to obtain standardized clinical data;

[0178] A pyramid feature fusion model for multimodal images, 30, is used to extract features from multi-scale images;

[0179] The DAFT clinical information fusion module 40 is used to fuse clinical information with image features from multiple MRI sequences.

[0180] After fusing clinical information and imaging features from multiple MRI sequences, the data is processed by the Region Generation Network 50 and the ROIAlign module 60 to output branches for automatic lesion detection, internal opening localization, perianal abscess / anal fistula classification, and pathogen classification.

[0181] This invention uses DAFT technology to fuse clinical information with multiple MRI sequences. It also includes multi-channel input of MRI sequences using T1WI, T2WI, and T2WI_FS, as well as a multi-task output method that simultaneously performs lesion detection (including perianal abscess, anal fistula, and internal opening), anal fistula classification, perianal abscess classification, and perianal abscess pathogen classification.

[0182] The specific construction of the pyramid feature fusion model 30 for multimodal images and the DAFT clinical information fusion module 40 is shown in Example 1, and will not be repeated here.

[0183] Based on the MRI deep learning framework, a DAFT method was used to integrate clinical information on the basis of a multimodal pyramid feature fusion network, and an intelligent image-assisted decision-making system was constructed to automatically detect and diagnose perianal abscesses and anal fistulas throughout the entire course of the disease. It covers multi-dimensional image assessment of anal fistula classification and internal opening location, perianal abscess classification and pathogen classification, with high diagnostic efficiency and homogeneity, which is conducive to accurate clinical decision-making.

[0184] Example 3

[0185] This embodiment provides an MRI device that is compatible with the decision support method of Embodiment 1, including the decision support system of Embodiment 2. It can be used as a standalone device or as an additional functional module loaded onto a conventional MRI instrument.

[0186] While specific embodiments of the invention have been described in detail by way of examples, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of the invention. Those skilled in the art should understand that modifications can be made to the above embodiments without departing from the scope and spirit of the invention. The scope of the invention is defined by the appended claims.

Claims

1. A medical image intelligent auxiliary decision method for perianal abscess and anal fistula, characterized in that, include: The original MRI sequence images were acquired, including T1-weighted imaging, T2-weighted imaging, and T2W1 fat-suppressed images. The original MRI sequence images were preprocessed to obtain preprocessed T1-weighted imaging, preprocessed T2-weighted imaging, and preprocessed T2W1 fat-suppressed images. Acquire clinical data, which includes at least: demographic characteristics, medical history, relevant laboratory tests, anoscopy, pathogen culture and detection results, pathological results, and MRI diagnostic results. Preprocess the clinical data to obtain standardized clinical data. A multi-task deep learning model is constructed, including: a feature extraction network layer, a region generation network, an output network layer, and a DAFT clinical information fusion module. This multi-task deep learning model is used for automatic lesion detection, internal opening localization, perianal abscess / anal fistula classification, and pathogen classification. The feature extraction network layer includes a deep neural network, Swin Transformer-Tiny, and a feature pyramid network, FPN. It has multi-label, multi-channel input and uses Swin Transformer-Tiny and FPN to obtain feature maps from the three modalities: preprocessed T1-weighted imaging, preprocessed T2-weighted imaging, and preprocessed T2W1 fat-suppressed imaging. ; The region generation network processes feature maps Generate candidate regions of interest to obtain candidate regions; The output network layer includes a lesion segmentation branch, an internal opening localization branch, and a perianal abscess / anal fistula classification branch. An ROI Align module is set between the region generation network and the output network layer to extract a fixed-size region feature map from the candidate region. The lesion segmentation branch is based on the output features of the ROI Align module, and uses the CBAM attention module to enhance the expression of spatial and channel information, and outputs a lesion segmentation mask; The inner opening localization branch uses the Coord Attention module to enhance the coordinate attention of possible inner opening positions in the candidate region and output the inner opening heatmap or localization coordinates. The classification branch uses the Squeeze-and-Excitation module to perform channel weighting on candidate regions or fusion features, and outputs the classification results of anal fistula and perianal abscess. The pathogen classification task combines the features after fusion of clinical information to perform a three-class classification output, resulting in automatic lesion detection, internal opening localization, perianal abscess / anal fistula classification, and pathogen classification using the multi-task deep learning model. The DAFT clinical information fusion module is used to dynamically adjust the feature fusion results of the standardized clinical data and input images, including: Global average pooling is performed on the region feature map output by the ROI Align module to obtain a length of The channel-level global feature vector is then used to obtain the channel-level global feature vector; this vector is then combined with the clinical information vector. The features are concatenated to form a joint feature representation; the joint feature is then compressed through a first fully connected layer to... The dimension, after ReLU activation, is mapped to a length of [length missing] through a second fully connected layer. The vector, where the first Weiwei ,back Weiwei Among them, only for Apply activation function ,wrong Apply activation, and ultimately utilize and Channel-by-channel modulation is performed on the original region feature map to obtain a new feature map that incorporates clinical information. .

2. The intelligent assisted decision-making method for perianal abscess and anal fistula according to claim 1, characterized in that, Preprocessing of the original MRI sequence images includes sequential format conversion, denoising, bias field correction, registration, resampling, intensity normalization, and region of interest cropping.

3. The intelligent assisted decision-making method for perianal abscess and anal fistula according to claim 1, characterized in that, Feature maps were obtained from the three modalities—preprocessed T1-weighted imaging, preprocessed T2-weighted imaging, and preprocessed T2W1 fat-suppressed imaging—using the deep neural network Swin Transformer-Tiny and the feature pyramid network FPN. ,include: The three modalities—preprocessed T1-weighted imaging, preprocessed T2-weighted imaging, and preprocessed T2W1 fat-suppressed imaging—are respectively input into the feature extraction network layer, and the input set is... for: ; in, For preprocessed T1-weighted imaging, For preprocessed T2-weighted imaging, This is a pre-processed T2W1 fat-suppressed image; After four stages of Swin Transformer-Tiny feature extraction, multi-scale feature maps of the three modalities at the four stages are obtained and output. Let the first stage be... The modal image in the first The feature maps output at each stage are: ； in, These correspond to preprocessed T1-weighted imaging, preprocessed T2-weighted imaging, and preprocessed T2W1 fat-suppressed imaging, respectively. This corresponds to stages 1 through 4; Indicates the first The mode in the th ... The Swing Transformer-Tiny feature extraction operation in each stage; In each stage, a feature fusion module is set up between different modalities to perform cross-modal feature fusion of multimodal features at the same level. The outputs of the three modes in each stage are as follows: , , The feature map after feature fusion is represented as: ； in, This represents a feature concatenation operation along the channel dimension. Indicates the first Multimodal feature map after fusion of each stage; After obtaining the feature fusion at each stage, a self-oriented feature pyramid is constructed using FNP. High-level feature maps are upsampled and then horizontally connected and added to adjacent low-level feature maps, forming a multi-scale feature representation that balances high-level semantic information and low-level spatial details. This includes setting the first... The features obtained after upsampling and lateral connection in each stage of the FPN path are: Then the first The final output features of each stage are represented as follows: ； in, This indicates element-wise addition. Indicates the first The feature maps output after each stage are fused by FPN are used as multi-scale inputs for subsequent task branches.

4. The medical imaging-assisted decision-making method for perianal abscess and anal fistula according to claim 1, characterized in that, The region generation network processes feature maps Generate candidate regions of interest, resulting in candidate regions including: Feature map output after FPN fusion A sliding window is used to pre-define several anchor boxes at each location. The classification score for each anchor box (whether it contains the target) and the regression offset from that anchor point to the true target box are predicted. The classification and regression are expressed as follows: ； ； in, This indicates the probability that the candidate box is the target region. This represents the regression results for the candidate bounding box coordinates.

5. The intelligent assisted decision-making method for perianal abscess and anal fistula according to claim 1, characterized in that, Extracting a fixed-size region feature map from the candidate region, the expression is: ； in, This represents a fixed-size feature map extracted from the candidate region. This indicates the coordinate information of the candidate region.

6. The intelligent assisted decision-making method for perianal abscess and anal fistula according to claim 1, characterized in that, An uncertainty-weighted strategy is used to jointly optimize the multi-task loss, assuming the lesion segmentation loss is... The internal positioning loss is Fractal loss is The total loss can then be expressed as: ； in, For the task The learnable uncertainty parameter.

7. The intelligent assisted decision-making method for perianal abscess and anal fistula according to claim 1, characterized in that, The DAFT clinical information fusion module is used to dynamically adjust the feature fusion results of the standardized clinical data and input images, including: Suppose that for the i-th segmentation instance in the dataset, its clinical information vector Represented as: ； Its ROI-level image feature map Represented as: ； in, Representing the channel index, the DAFT feature vector fusion module uses an auxiliary network. Scaling factor is generated based on image features and clinical information. and offset Then, a channel-by-channel affine transformation is performed on the original feature map: ； ； ； in, and This is a mapping function modeled by an auxiliary neural network.

8. A medical imaging-based intelligent decision-making system for perianal abscess and anal fistula, characterized in that, The decision support method applied to any one of claims 1 to 7 includes: an image preprocessing module, a clinical data preprocessing module, a pyramid feature fusion model for multimodal images, a DAFT clinical information fusion module, a region generation network, and an ROI Align module, wherein... The image preprocessing module is used to preprocess the acquired MRI sequence raw images to obtain preprocessed T1-weighted images, preprocessed T2-weighted images, and preprocessed T2W1 fat-suppressed images. The clinical data preprocessing module is used to preprocess the clinical data to obtain standardized clinical data. The pyramid feature fusion model for multimodal images is used to extract multi-scale image features; The DAFT clinical information fusion module is used to fuse clinical information with image features from multiple MRI sequences. After fusing clinical information and imaging features from multiple MRI sequences, the data is processed by the region generation network and the ROIAlign module to output branches for automatic lesion detection, internal opening localization, perianal abscess / anal fistula classification, and pathogen classification.

9. An MRI device, characterized in that, It includes the decision support system as described in claim 8, and performs the decision support method as described in any one of claims 1 to 7.

Images