Pneumonia pathological image detection method and system based on frequency domain analysis and dynamic convolution

Abstract

The application discloses a pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution and belongs to the technical field of computer vision, intelligent medical image analysis and deep learning target detection. The method comprises the following steps: a lung inflammation focus spectrum prototype library is constructed, a learnable annulus-sector spectrum mask and an amplitude-phase coupling gate are used to selectively enhance the focus related spectrum, bidirectional gate space-frequency fusion guided by a candidate focus is used to realize mutual verification of spatial morphology abnormalities and spectrum abnormalities, and a pathological perception dynamic convolution generated by spectrum prototype modulation is used to generate a detection convolution kernel adaptive to the type, scale and boundary state of the candidate focus, so that the detection accuracy and robustness of ground glass shadows, early small patchy foci, real shadow edges and multiple focus lesions are improved. The application further discloses a pneumonia pathological image detection system based on frequency domain analysis and dynamic convolution, an electronic device and a storage medium for realizing the above method.

Description

Technical Field

[0001] This invention relates to a method and system for detecting pneumonia pathological images, specifically a method and system for detecting pneumonia pathological images based on frequency domain analysis and dynamic convolution, belonging to the fields of computer vision, intelligent medical image analysis, and deep learning target detection technology. Background Technology

[0002] Pneumonia lesions typically present in various forms on chest X-rays and CT scans, including ground-glass opacities, patchy consolidation, linear opacities, interstitial thickening, or multifocal small patches. These lesions not only possess spatial morphological, boundary, and locational characteristics but also frequency domain features such as energy distribution, texture directionality, and local phase consistency. Ground-glass opacities and early inflammatory exudative lesions, in particular, show relatively small grayscale differences compared to normal lung texture, vascular texture, and rib projections; relying solely on spatial convolution can easily lead to missed detections. Conversely, enhancing only high-frequency information may misinterpret rib edges, vascular bifurcation, or noise as lesions.

[0003] Existing computer-aided diagnostic methods attempt to improve detection or segmentation accuracy using techniques such as convolutional neural networks, feature pyramids, attention mechanisms, frequency domain transformation, and dynamic convolution. However, pneumonia lesions are not limited to fixed frequency bands: ground-glass opacities typically exhibit fine-grained texture perturbations in the mid-to-high frequencies, solid opacities tend to show low-to-mid-frequency energy accumulation, and small patchy inflammatory lesions rely on local spatial boundaries and abnormal peaks in the frequency domain. Therefore, simply splicing frequency domain features as ordinary branches with spatial features, or using only general channels / spatial attention to enhance the spectrum, is insufficient for accurate detection of pneumonia pathological images. Summary of the Invention

[0004] This invention aims to provide a method and system for detecting pneumonia pathological images based on frequency domain analysis and dynamic convolution, addressing the problems of insufficient generalization ability of frequency domain enhancement, coarse granularity of spatial-frequency fusion, and lack of pathological modulation in dynamic convolution in existing technologies. This invention constructs a spectrum prototype library of pneumonia lesions, utilizes learnable annular-sector spectrum masks and amplitude-phase coupled gating to selectively enhance the spectrum related to lesions; achieves mutual verification of spatial morphological anomalies and spectral anomalies through bidirectional gating spatial-frequency fusion guided by candidate lesions; and generates detection convolution kernels adapted to the type, scale, and boundary state of candidate lesions through pathologically perceptive dynamic convolution modulated by the spectrum prototype, thereby improving the detection accuracy and robustness of ground-glass opacities, early small patchy lesions, solid edges, and multifocal lesions.

[0005] To achieve the above objectives, this invention provides a method for detecting pneumonia pathological images based on frequency domain analysis and dynamic convolution, comprising the following steps:

[0006] Step 1: Acquire pneumonia pathological images, preprocess the pneumonia pathological images and determine the lung field constraint region, and establish a learnable lesion spectrum prototype based on the lung field constraint region;

[0007] Step 2: Input the preprocessed pneumonia pathological image and lesion spectrum prototype into the frequency domain-spatial domain dual-branch encoder for processing to obtain spatial features and the frequency domain enhancement feature map, spectrum anomaly map and spectrum prototype response vector output by the frequency domain enhancement module guided by the lesion spectrum prototype;

[0008] Step 3: Using the candidate lesion to guide the bidirectional gated space-frequency fusion module, the spatial domain features and the frequency domain enhanced feature map are fused together with the spectrum anomaly map to obtain mutual verification candidate features;

[0009] Step 4: Input the mutual verification candidate features and the spectrum prototype response vector into the spectrum prototype modulation pathological perception dynamic convolution module for processing. Use the generated dynamic detection convolution kernel to obtain the dynamic convolution response, and add the dynamic convolution response to the mutual verification candidate features by residual addition to output the candidate detection features.

[0010] Step 5: Process the candidate detection features using the decoupled detection head and output the final pneumonia lesion detection results.

[0011] Furthermore, in step 1, the spectrum prototype includes the amplitude spectrum, phase consistency spectrum, and direction spectrum of the lesion region and the normal lung texture region.

[0012] Furthermore, in step 2, the process of the lesion spectrum prototype-guided frequency domain enhancement module processing features includes:

[0013] A two-dimensional fast Fourier transform is performed on the input feature map to obtain the complex spectrum, which is then separated into the amplitude spectrum and the phase spectrum.

[0014] The spectrum region is divided by a learnable annular-sector spectral mask, and each region is matched with the lesion spectrum prototype to obtain the matching result.

[0015] Based on the matching results, amplitude enhancement gating and phase fidelity gating are generated. The amplitude spectrum and phase spectrum are processed by amplitude enhancement gating and phase fidelity gating respectively to obtain the enhanced amplitude spectrum and the constrained phase spectrum.

[0016] The enhanced amplitude spectrum and the constrained phase spectrum are recombined into a complex spectrum, which is then subjected to inverse Fourier transform to obtain a frequency domain enhanced feature map. At the same time, a spectral anomaly map and a spectral prototype response vector are output.

[0017] Further, step 3 includes:

[0018] The spatial features are processed to generate an initial candidate heatmap; the spectral anomaly map output by the frequency domain enhancement module guided by the lesion spectrum prototype in step 2 is reused, and a scale reliability map is generated according to the coding level of the spatial features and the size of the candidate box.

[0019] The space-guided frequency-domain gating branch receives spatial features, frequency-domain enhanced feature maps, candidate heatmaps, spectral anomaly maps, and scale reliability maps to generate a space-guided frequency-domain gating. This gating is used to weight the frequency-domain enhanced feature map and add it to the spatial feature residual to obtain the first branch feature. The frequency-domain-guided spatial gating branch generates a frequency-domain-guided spatial gating based on the spectral anomaly map and the lesion spectral prototype response vector. This gating is used to weight the spatial features and add it to the frequency-domain enhanced feature map residual to obtain the second branch feature.

[0020] The first branch feature and the second branch feature are added together to obtain a bidirectional fusion base; the first branch feature and the second branch feature are subjected to mutual verification consistency judgment to generate a consistency weight; the consistency weight is applied to the bidirectional fusion base to generate a lesion enhancement term; the supplementary amount of the consistency weight is applied to the bidirectional fusion base to generate a background suppression term; the lesion enhancement term is superimposed on the bidirectional fusion base and the background suppression term is subtracted to obtain mutual verification candidate features.

[0021] Further, step 4 includes:

[0022] First, candidate morphological vectors are extracted from the spatial features, spectral response vectors are extracted from the spectral prototype response vectors output by the lesion spectral prototype-guided frequency domain enhancement module, and scale vectors are generated according to the current coding level and candidate box size.

[0023] Then, the candidate morphological vector, spectral response vector and scale vector are concatenated and input into the kernel modulation network for processing to generate basic convolutional kernel combination weights. The basic convolutional kernel combination weights include block convolutional kernel weights, directional convolutional kernel weights, dilated convolutional kernel weights and high-resolution small lesion convolutional kernel weights, and the sum of the above weights after normalization is one.

[0024] Next, the preset basic convolutional kernel group is adaptively weighted and summed using the weights of the basic convolutional kernel combination to generate a dynamic detection convolutional kernel;

[0025] Finally, the candidate features for mutual verification are convolved using a dynamic detection convolution kernel to obtain a dynamic convolution response. The dynamic convolution response is then added to the residual of the candidate features for mutual verification to obtain the candidate detection features.

[0026] Furthermore, the present invention also includes: end-to-end training using a joint loss function, wherein the joint loss function includes detection loss, spectral prototype constraint loss, phase fidelity loss, spatial-frequency mutual verification consistency loss, and dynamic kernel smoothing loss; wherein the detection loss includes classification loss, bounding box regression loss, and confidence loss.

[0027] A pneumonia pathological image detection system based on frequency domain analysis and dynamic convolution includes:

[0028] The lung image acquisition and lung field constraint preprocessing unit is used to acquire pneumonia pathological images, preprocess the pneumonia pathological images, and determine the lung field constraint area.

[0029] The suspected lesion spectrum prototype establishment unit is used to establish a learnable lesion spectrum prototype based on the lung field constraint region.

[0030] The lesion spectrum prototype-guided frequency domain enhancement unit is used to input the preprocessed pneumonia pathological image and lesion spectrum prototype into the frequency domain-spatial domain dual-branch encoder for processing, to obtain spatial features and the frequency domain enhancement feature map, spectrum anomaly map and spectrum prototype response vector output by the lesion spectrum prototype-guided frequency domain enhancement module;

[0031] The candidate suspected lesion guides a bidirectional gated space-frequency fusion unit, which is used to fuse spatial domain features and frequency domain enhanced feature maps using a bidirectional gated space-frequency fusion module and a spectral anomaly map to obtain mutually verifiable candidate features;

[0032] The Spectrum Prototype Modulation Pathological Perception Dynamic Convolution Unit is used to input the mutual evidence candidate features and the spectral prototype response vector into the Spectrum Prototype Modulation Pathological Perception Dynamic Convolution for processing. The generated dynamic detection convolution kernel is used to obtain the dynamic convolution response, and the dynamic convolution response is added to the mutual evidence candidate feature residual to output the candidate detection feature.

[0033] The decoupled detection output unit is used to process candidate detection features using the decoupled detection head and output the final pneumonia lesion detection result;

[0034] The joint training unit is used for joint training using detection loss, spectral prototype constraint loss, phase fidelity loss, spatial-frequency mutual verification consistency loss, and dynamic kernel smoothing loss.

[0035] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0036] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1. This invention establishes spectral prototypes of ground-glass opacities, consolidations, linear opacities, and normal lung textures by referencing the same lung field background. This ensures that the frequency domain enhancement weights, candidate fusion weights, and detection convolution kernels are all driven by the spectral prototype response of the same candidate lesion, forming an inseparable technical chain of lesion spectral prototype identification, candidate spatial-frequency verification, and dynamic detection convolution kernel modulation. This chain allows the dynamic convolution kernel to adaptively adjust according to the pathological spectrum type, scale, and boundary ambiguity of the candidate lesion, thereby reducing the missed detection of low-contrast lesions.

[0039] 2. This invention guides the frequency domain enhancement module using a learnable annular-sector spectral mask to jointly divide radial frequencies and directional angles based on the lesion spectral prototype. In addition to amplitude enhancement gating, phase fidelity gating is introduced to limit the enhancement target to frequency-directional regions that match the spectral prototype of pneumonia lesions and significantly deviate from the normal background of the same lung field. This processing method establishes a pathological constraint relationship between spectral amplitude, directional energy, phase gradient, and lung field background reference, which can enhance the weak texture abnormalities of ground-glass opacities and early small patchy lesions, while suppressing false enhancement of rib edges, vascular bifurcation, and high-frequency noise components.

[0040] 3. This invention generates initial candidate heatmaps, spectral anomaly maps, and scale reliability maps at each detection scale through a candidate lesion-guided bidirectional gating space-frequency fusion module, and performs spatial-guided frequency-domain gating and frequency-domain-guided spatial gating. The two gating branches respectively form enhancement features with residuals, which are added together to form the bidirectional fusion base. A consistency weight is generated through mutual verification; this weight acts on the bidirectional fusion base to form a lesion enhancement term, and its complement acts on the same base to form a background suppression term. When spatial candidates and pneumonia spectrum anomalies mutually corroborate each other, the lesion enhancement term dominates; when they are inconsistent, the background suppression term dominates. This mechanism effectively reduces false positives such as vascular crossings, rib projections, and lung texture aggregation, rather than simply obtaining feature enhancements after general space-frequency fusion. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the overall network architecture provided by the present invention;

[0042] Figure 2 This is a schematic diagram of the structure of the lesion spectrum prototype-guided frequency domain enhancement module provided by the present invention;

[0043] Figure 3 A schematic diagram of the structure of the spectrum prototype modulation pathological perception dynamic convolution module provided by the present invention;

[0044] Figure 4 A schematic diagram of the candidate lesion-guided bidirectional gated space-frequency fusion module provided by the present invention;

[0045] Figure 5 This is a schematic diagram of the decoupling detection head provided by the present invention. Detailed Implementation

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0047] Example 1:

[0048] This embodiment provides a pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution, which includes the following steps:

[0049] The overall network execution order in this embodiment is as follows: Figure 1 As shown, the process involves first performing frequency domain enhancement guided by the lesion spectrum prototype, then performing bidirectional gated space-frequency fusion guided by candidate lesions, and finally performing pathological perception dynamic convolution and decoupled detection output based on spectrum prototype modulation.

[0050] Step 1: Data preparation, lung field constraint preprocessing, and spectral prototype initialization.

[0051] Acquire public or private pneumonia pathology image datasets, including chest X-rays, chest CT slices, or 2D images obtained from 3D CT resampling. After image size and grayscale normalization, the initial lung field extent is first obtained through coarse lung field segmentation, followed by morphological correction or a lightweight segmentation network to estimate a binary lung field mask; both determine the lung field constraint region, which limits the effective range of subsequent spectral statistics, candidate region generation, and background reference sampling. During the training phase, lesion candidate regions are obtained by cropping within the lung field constraint region using bounding boxes or bounding masks; the normal lung texture reference region in the same image is obtained by excluding lesion bounding regions, high-response areas at lung field boundaries, and strong edge regions of major blood vessels / rib projections within the same lung field. During the inference phase, lesion candidate regions are given by the initial candidate heatmap or detection candidate boxes, while the normal lung texture reference region is still obtained from the low lesion response region under the lung field mask constraint.

[0052] For each type of pneumonia lesion, during the training phase, the radial energy, directional energy, and phase gradient consistency of the amplitude spectrum of the lesion candidate region are statistically analyzed. Similarly, the normal lung texture reference region in the same image is also statistically analyzed, thereby initializing the spectral prototypes of ground-glass opacities, consolidations, linear opacities, and normal lung textures. These lesion spectral prototypes are updated as learnable parameters in subsequent training. Specifically, the lesion candidate region is used to form the lesion spectral prototype, and the normal lung texture reference region is used to calculate the background spectral deviation, suppress false enhancement of vascular / rib textures, and provide a background baseline for the spectral anomaly map.

[0053] The input lung medical images, lung field masks, candidate suspected lesion region sets, scale features, and spectral prototype sets are defined as follows:

[0054] ,

[0055] ,

[0056] Where X represents the input lung medical image, and H, W, and C represent its height, width, and number of channels, respectively; R represents a lung field mask with a value of 0 or 1, used to constrain spectral statistics, candidate generation, and background reference sampling regions; R represents the region of N candidate suspected lesions. The set constituted; P represents the set of spectrum prototypes, where , , , These represent the spectral prototypes of ground-glass opacities, consolidation opacities, linear opacities, and normal lung textures, respectively. P is used for spectral similarity matching and gating generation in step 2, forms the spectral prototype response vector in step 3 to participate in space-frequency mutual verification, is used to modulate the dynamic convolution kernel in step 4, and is continuously updated through spectral prototype constraint loss during joint training.

[0057] Step 2: Input the preprocessed image into a frequency-spatial dual-branch encoder. The spatial branch outputs a multi-scale spatial feature map, while the frequency branch outputs lesion-related spectral enhancement features, spectral anomaly maps, and spectral prototype response vectors at the corresponding scales.

[0058] Spatial domain branch: Images are processed using convolutional backbone networks or Transformer backbone networks to extract spatial structural features, resulting in spatial feature map sequences at different scales. The spatial domain branch primarily encodes the anatomical location of the lung field, candidate lesion morphology, boundary continuity, and relative relationship with structures such as blood vessels and ribs.

[0059] Frequency domain branch: The shallow feature map of the image or spatial domain branch is input into the lesion spectral prototype to guide the frequency domain enhancement module, namely the Pathological Frequency Enhancement Module (PFEM), hereinafter referred to as the frequency domain enhancement module. This frequency domain enhancement module uses the pneumonia lesion spectral prototype and the difference in background spectrum within the same lung field as the core to generate frequency domain enhancement weights, rather than simply calculating channel attention and spatial attention in parallel.

[0060] like Figure 2 As shown, the specific processing procedure of the frequency domain enhancement module is as follows:

[0061] 1. Amplitude-Phase Spectrum Transform: A two-dimensional Fast Fourier Transform is performed on the input feature map to obtain a complex spectrum, separating the amplitude spectrum and the phase spectrum. The amplitude spectrum represents the texture energy distribution, while the phase spectrum preserves the spatial location and boundary morphology of the lesions, represented as:

[0062] ,

[0063] in, This represents the spatial branch feature map of the l-th layer; Indicates candidate region The spectral mask mapped to the l-th layer; This indicates element-wise multiplication; Represents a two-dimensional Fast Fourier Transform; Represents the complex spectrum of the candidate region at layer l; and They represent respectively by The amplitude spectrum and phase spectrum obtained by separation; Indicates taking the modulus of a complex number. This indicates taking the complex phase.

[0064] 2. Learnable annular-sector mask: Multiple radial annular bands and directional sectors are established in the spectral plane. The annular band radius, sector angle, and bandwidth are learned by the network and constrained by the normal background spectrum within the lung field. This results in a finer frequency-directional decomposition than fixed high-pass / low-pass filtering, capturing the fine-grained mid-to-high frequency perturbations of ground-glass opacities, the low-to-mid frequency blocky energy of solid opacities, and the directional texture of linear opacities.

[0065] Furthermore, K learnable ring-sector masks are established in the spectral plane to obtain spectral descriptors for candidate regions, and the corresponding spectral prototypes are updated based on the lesion labels:

[0066] ,

[0067] ,

[0068] in, This represents the k-th learnable annular-sector mask; This indicates the extraction of amplitude mean, amplitude variance, phase gradient, and directional energy statistics from the complex spectrum within the mask; This indicates that the above statistics are coded using features; Indicates candidate region Spectral descriptor within the k-th mask; Represents the spectral prototype of category c; Indicates the prototype momentum update coefficient; Indicates the satisfaction Candidate region spectral descriptor Find the mean; Indicates candidate region Lesion category labels.

[0069] 3. Lesion spectrum prototype matching: Prototype similarity is defined as follows: .

[0070] in, Indicates candidate region Spectral descriptor within the k-th mask With category c prototype The similarity between them; cos represents the cosine similarity; and represents the learnable projection matrices acting on the spectral descriptor and the spectral prototype, respectively. Here, the vector similarity is calculated after the matrix is ​​projected.

[0071] 4. Amplitude Enhancement and Phase Preservation: An amplitude enhancement gating is generated based on the spectral prototype matching results, while a phase preservation gating is generated based on the candidate region's central phase, boundary phase gradient, and lung field reference phase. The amplitude enhancement gating selectively enhances lesion-related frequency components, while the phase preservation gating limits phase drift during frequency domain reconstruction, ensuring that the enhanced features remain aligned with the original lesion location.

[0072] Amplitude enhancement gating and phase fidelity gating are defined as follows:

[0073] ,

[0074] ,

[0075] in, The spectral deviation of the candidate region relative to the normal lung texture reference region in the same image can be composed of the difference between the spectral descriptor of the candidate region and the normal lung texture reference descriptor, the L2 distance, and the similarity difference relative to the background prototype. The normalized prior distance representing the candidate center to the lung field boundary and the strong edge texture of major blood vessels / ribs can be calculated jointly by the lung field mask distance transformation, the edge response map of blood vessels / bony structures, and the coordinates of the candidate center.

[0076] in, Indicates amplitude enhancement gating. Indicates phase fidelity gating, This represents the Sigmoid activation function; , , , These represent the similarity between the candidate region and the ground glass shadow, real shadow, stripe shadow, and background prototype, respectively. and For learnable parameters, and These represent the phase gradient of the candidate center and the phase gradient of the candidate boundary, respectively.

[0077] 5. Inverse Transform Output: The enhanced amplitude spectrum and the constrained phase spectrum are recombine into a complex spectrum. An inverse Fourier transform is then performed to obtain the frequency domain enhanced feature map. Simultaneously, a spectral anomaly map and a spectral prototype response vector are output for use by the subsequent space-frequency fusion module and dynamic convolution module. The enhanced spectrum and its inverse transform output are:

[0078] ,

[0079] ,

[0080] in, Represents the enhanced complex spectrum; , These represent the amplitude enhancement intensity coefficient and the phase constraint intensity coefficient, respectively, initially taken in a preferred reproducible setting. , And can be based on the validation set in Adjust within the range; Indicates the constrained phase correction amount; Represents the imaginary unit; Indicates complex exponentiation; Indicates the two-dimensional inverse Fourier transform The reconstructed frequency domain enhanced feature map.

[0081] Spectral anomaly graph In step 2, the frequency domain enhancement features The maximum response and background spectral deviation of the prototype similarity of each lesion category are concatenated and then subjected to convolution and... Activation generation; spectral prototype response vector The candidate region is formed by pooling along the mask dimension based on the category similarity of each annulus-sector. Its components correspond sequentially to the prototype responses of ground-glass opacity, consolidation, linear opacity, and normal lung texture. Step 3 involves direct reuse. and Spectral anomaly maps will no longer be generated repeatedly.

[0082] Step 3: Using the candidate lesion to guide the bidirectional gated space-frequency fusion module, and combining the spectral anomaly map to perform feature fusion on the spatial domain features and frequency domain enhanced feature map, to obtain mutual verification candidate features.

[0083] To achieve mutual verification between spatial morphological evidence and spectral anomaly evidence, this invention constructs a candidate lesion-guided bidirectional gated frequency fusion module (LBG-FFM) at each feature scale of the encoder. This module first generates an initial candidate heatmap based on spatial features, directly reuses the spectral anomaly map output by the frequency domain enhancement module in step 2, and generates a scale reliability map based on the coding level of the current feature scale and the candidate box size. The spatially guided frequency domain gating branch receives spatial features, the frequency domain enhancement feature map, the candidate heatmap, the spectral anomaly map, and the scale reliability map, generating a spatially guided frequency domain gating. This gating is used to weight the frequency domain enhancement feature map and added to the spatial feature residual to obtain the first branch feature. The frequency domain-guided spatial gating branch generates a frequency domain-guided spatial gating based on the spectral anomaly map and the lesion spectral prototype response vector. This gating is used to weight the spatial features and added to the frequency domain enhancement feature map residual to obtain the second branch feature.

[0084] The structure of the candidate lesion-guided bidirectional gated space-frequency fusion module is as follows: Figure 4 As shown. Figure 4 The upper branch corresponds to spatially guided frequency domain gating and the first branch residual output, while the lower branch corresponds to frequency domain guided spatial gating and the second branch residual output. The features from both branches are input to the mutual evidence consistency judgment and added together to form a bidirectional fusion base. The mutual evidence consistency weight is applied to this base to form a lesion enhancement term, and its complement is applied to the same base to form a background suppression term. Finally, the lesion enhancement term is superimposed on the base, and the background suppression term is subtracted to output the mutual evidence candidate features. The input terms, two residual addition nodes, mutual evidence consistency judgment, lesion enhancement, and background suppression paths in the following formulas are all related to... Figure 4 The connection relationships correspond.

[0085] The spatial candidate heatmap, spectral anomaly map, and scale reliability map are defined as follows:

[0086]

[0087]

[0088]

[0089] in, Indicates spatial characteristics Convolution and Activation function The generated initial candidate heatmap; This indicates that the spectral anomaly map generated by the lesion spectrum prototype guided by the frequency domain enhancement module in step 2 and directly reused in this step has been obtained. This represents the spectral anomaly diagram output by the module at layer l; Represents the reliability diagram based on scale; , These represent the width and height of the candidate box, respectively. This represents the learnable parameters used to generate the scale reliability map.

[0090] Based on this, the two gating mechanisms, corresponding branch residual features, mutual evidence consistency weights, bidirectional fusion basis, lesion enhancement term, background suppression term, and mutual evidence fusion features are as follows:

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100] in, Indicates spatial characteristics Frequency domain enhanced feature map Candidate heatmap Spectrum anomaly diagram and scale reliability diagram After splicing, convolution and The resulting spatially guided frequency domain gating is activated; Indicates the use of right Weighted and The first branch feature is obtained by adding the residuals. Indicated by the spectral anomaly plot and spectrum prototype response vector After splicing, convolution and The frequency domain-guided spatial gating obtained upon activation; Indicates the use of right Weighted and The second branch feature is obtained by adding the residuals. This represents the mutual verification consistency weight generated based on the features of the two branches; This represents a bidirectional fusion basis formed by adding features from two branches; This indicates that the consistency weight is applied to the enhancement term of the lesion obtained by the bidirectional fusion substrate; This represents the complementary portion of the bidirectional fusion matrix after deducting the lesion enhancement term, i.e., the background suppression term corresponding to the consistency weight complement. This represents the mutual verification and fusion feature of the l-th layer. and These represent the non-negative weights of the lesion enhancement term and the background suppression term, respectively. In a preferred setting, both are set to 1, and can be adjusted based on the validation set; square brackets indicate feature splicing relationships.

[0101] The candidate lesion-guided bidirectional gated spatial-frequency fusion module first generates bidirectional enhancement features with residuals through two gated branches, and then adds the features of the two branches to form a bidirectional fusion base. Subsequently, a mutual evidence consistency weight is generated based on the consistency between the two branch features. This weight acts on the bidirectional fusion base to form a lesion enhancement term, and its complement acts on the same base to form a background suppression term. Finally, the lesion enhancement term is superimposed on the bidirectional fusion base, and the background suppression term is subtracted. When both the spatial candidate and the spectral anomaly point to the same region, the consistency weight increases, and the lesion enhancement term dominates; when the spatial morphology and spectral evidence are inconsistent, the consistency weight decreases, and the background suppression term dominates, thereby suppressing background responses such as vascular crossings, rib projections, and lung texture aggregation. This processing corresponds to... Figure 4 The two additive nodes, mutual verification consistency judgment, lesion enhancement and background suppression dual paths.

[0102] The fused multi-scale features are aggregated across different coding levels of the backbone network, and the spectral prototype response vector of the same lesion is shared between adjacent coding levels. For the same lesion whose size varies across different coding levels, the network maintains the continuous response of the dynamic convolutional kernel at different coding levels through cross-scale dynamic kernel smoothing constraints, thereby reducing the missed detection of small lesions at the level switching points.

[0103] Step 4: In different coding levels of the backbone network, a Spectral-Prototype Modulated Pathology-Aware Dynamic Convolution (SP-PADC) module is used to replace part of the standard convolution, hereinafter referred to as the dynamic convolution module. For each candidate unit or candidate box, the dynamic convolution module first extracts the candidate morphological vector from the spatial domain features, extracts the spectral response vector from the spectral prototype response vector output by the frequency domain enhancement module, and generates a scale vector according to the coding level of the current feature and the size of the candidate box; then, the three are concatenated and input into the kernel modulation network to generate the basic convolution kernel combination weights.

[0104] The four types of basic convolutional kernel weight generation and dynamic kernel combination structure of the dynamic convolution module are as follows: Figure 3 As shown. Figure 3 In the process, the four types of basic convolutional kernels are weighted and summed according to the weights of the kernel modulation network output to form a dynamic detection convolutional kernel. The dynamic detection convolutional kernel is applied to the mutual verification candidate features obtained in step 3, and its convolutional response is added to the residual of the mutual verification candidate features to obtain the candidate detection features.

[0105] Furthermore, for each candidate region, the candidate morphological vector, spectral response vector, and scale vector are input into the kernel modulation network. The kernel modulation network outputs a set of normalized weights for the basic convolutional kernels, each component of which corresponds to one of the four types of basic convolutional kernels described below, as follows:

[0106]

[0107]

[0108] in, , , Representing candidate regions Candidate morphological vectors, spectral response vectors, and scale vectors; This represents the combined weights of the basic convolutional kernels, which have four components. , , , These correspond to block convolution kernels, directional convolution kernels, dilated convolution kernels, and high-resolution small lesion convolution kernels, respectively, and are processed by... The sum after normalization is one; This represents a multilayer perceptron used to generate the combined weights of four basic convolutional kernels. This represents the normalized exponential function.

[0109] The dynamic convolution module pre-configures a set of basic convolution kernels, including: a large receptive field block convolution kernel suitable for low-to-mid-frequency block structures (such as real lesions); a directional convolution kernel for mid-to-high-frequency directional fine textures (such as streaks); a dilated convolution kernel for regions with blurred boundaries (such as the edges of ground-glass opacities); and a high-resolution small lesion convolution kernel suitable for tiny lesions. The final dynamic detection convolution kernel is generated by adaptively weighting and summing the above four types of basic convolution kernels according to the combined weights of the basic convolution kernels.

[0110] The resulting dynamic detection convolutional kernel and candidate detection features are as follows:

[0111]

[0112]

[0113] in, This represents the dynamic detection convolution kernel obtained by weighted summation of four basic convolution kernels; , , , These represent block convolution kernels, directional convolution kernels, dilated convolution kernels, and high-resolution small lesion convolution kernels, respectively. This represents the layer l mutual verification fusion feature output from step 3. The feature of the i-th candidate region extracted from it; This represents the convolution operation; Represents the dynamic convolution response; This represents the candidate detection features obtained by adding the dynamic convolutional response to the residual features of the input candidate region. Figure 3 The adder node at the end.

[0114] Unlike methods that rely solely on low-level features or low-frequency information to generate decoder convolutional kernels, the convolutional kernels generated in this step modulate the pathological features of the target detection candidate region. This enables targeted enhancement of candidate features before classification, localization, and confidence prediction, thereby improving the response to low-contrast lesions.

[0115] Step 5: Decoupling detection and result output.

[0116] The fused features processed by SP-PADC and LBG-FFM are input into the decoupled detection head. The detection head includes a classification branch, a regression branch, a confidence branch, and an optional lesion spectrum type auxiliary branch. The lesion spectrum type auxiliary branch is used to predict whether candidate regions are closer to ground-glass opacities, consolidations, linear opacities, or normal lung texture prototypes, providing interpretable spectral evidence for the classification branch.

[0117] Decoupling detection head and lesion spectrum type auxiliary branch structure such as Figure 5 As shown.

[0118] Classification branch: Predicts whether each candidate box is a pneumonia lesion and its possible subtypes, while receiving the lesion spectrum prototype response vector as auxiliary input to improve the ability to identify low-contrast ground-glass opacities and early patchy opacities.

[0119] Regression branch: Predicts the center point, width, height, and boundary offset of the candidate bounding box. The regression branch also utilizes the boundary phase gradient features of the phase-fidelity gated output to reduce the localization error caused by lesion position drift after frequency domain enhancement.

[0120] Confidence branch: predicts the probability that the candidate box contains the real lesion, and uses the consistency of spatial and frequency verification as a confidence correction factor to give higher confidence to candidate boxes that are valid in both spatial morphology and spectral anomalies.

[0121] Through threshold filtering, spectral cross-verification confidence correction, and non-maximum suppression, the final output includes the location, size, category confidence of pneumonia lesions, and optional spectral prototype interpretation results.

[0122] Step 6: Model training.

[0123] This invention employs an end-to-end approach to train the entire detection network. The joint loss function used includes detection loss, spectral prototype constraint loss, phase fidelity loss, space-frequency mutual verification consistency loss, and dynamic kernel smoothing loss, which can be expressed as:

[0124] ,

[0125] ,

[0126] ,

[0127] ,

[0128] ,

[0129] ,

[0130] in, Indicates the total loss; Indicates detection loss; , and These represent the classification loss, bounding box regression loss, and confidence loss, respectively. This indicates a loss of spectral contrast in the lesion. Indicates the temperature coefficient; This indicates phase fidelity loss; Represents the gradient operator; and These represent the phase spectra of the candidate region before and after enhancement, respectively; and These represent the phase gradients before and after enhancement, respectively; This indicates the extraction of the phase of the candidate center region; It represents the first norm. This represents the consistency loss of space-frequency mutual verification, where This represents the overlap coefficient loss. Represents the binary cross-entropy loss. and These represent the surveillance mask for the lesion region and the target mask, respectively. This represents the cross-scale dynamic kernel smoothing loss. and This represents the dynamic convolution kernel for the same candidate region at adjacent coding levels. Represents the square of the L2 norm; , , , , and These represent the weights of the corresponding loss terms.

[0131] Through the aforementioned training objectives, the model not only learns the boundaries and locations of lesions in the spatial domain, but also grasps the spectral anomalies, phase stability, and candidate-level spatial-frequency cross-verification relationships of the lesions relative to the background of the same lung field. Thus, the frequency domain enhancement module, the spatial-frequency fusion module, and the dynamic convolution module constitute an interdependent technical closed loop: the frequency domain enhancement module provides spectral evidence of the lesions, the spatial-frequency fusion module verifies the consistency between spatial candidates and spectral evidence, and the dynamic convolution module generates dynamic convolution kernels adapted to candidate lesions based on the verified spectral response. This closed-loop structure is difficult to replace simply with fixed high- and low-frequency filtering, general parallel attention, channel recalibration modules, or general spatial-frequency fusion methods.

[0132] Regarding training details, input images are uniformly resampled to a preset size; X-ray images undergo grayscale normalization and lung field cropping, while CT images undergo lung window truncation, slice thickness normalization, and 2D slicing or 3D thin-slice projection; data augmentation methods include random rotation, scale perturbation, contrast perturbation, artifact simulation, and slight spectral perturbation, but spectral perturbation must not disrupt the phase fidelity label. The optimizer can be AdamW or stochastic gradient descent with momentum; the initial learning rate, learning rate decay strategy, and number of training epochs are determined based on the experimental settings.

[0133] In a preferred embodiment, the classification loss weight is fixed at 1, and the bounding box regression loss weight is... Set to 1.0 for confidence loss weights. Set to 1.0, for the spectral prototype constraint loss weight. Set to 0.2, phase fidelity loss weight Set to 0.1, weight for consistency loss in space-frequency mutual verification. Set the weight to 0.3 for dynamic kernel smoothing loss. Take 0.05, temperature coefficient The weight is set to 0.07; the above weights can be adjusted based on the validation set while maintaining the dominance of the main detection loss. During training, the spectrum prototype can be frozen for several rounds of updates to stabilize the lung field background reference, and then the spectrum prototype, frequency domain enhancement module, space-frequency fusion module, dynamic convolution module, and detection head can be updated jointly; during the inference phase, only the forward detection process is retained, and no annotation information is required.

[0134] To facilitate those skilled in the art in implementing the present invention and conducting subsequent effectiveness verification, this embodiment provides the following experimental setup and evaluation scheme.

[0135] I. Experimental Data and Preprocessing.

[0136] Validation can be performed using publicly available chest X-ray datasets of pneumonia with lesion bounding boxes as the primary data source. If using publicly available RSNA pneumonia detection datasets, training, validation, and test sets should be divided according to patient dimensions to avoid images of the same patient appearing across sets. Input images can be uniformly resampled to 1024×1024 pixels, and grayscale normalization and lung field cropping should be performed. The sample size, data partitioning ratio, and exclusion criteria are determined based on the sample composition of the publicly available dataset used, ensuring that images of the same patient do not appear across sets.

[0137] II. Experimental Environment.

[0138] In a preferred embodiment, model training is performed using a deep learning workstation equipped with a graphics processing unit (GPU). The training framework is PyTorch, the optimizer is AdamW, the batch size is set to 8, and the initial learning rate is set to... The cosine annealing strategy was used to train the system for 100 rounds.

[0139] III. Evaluation Indicators.

[0140] Detection performance can be evaluated using mean precision, recall, accuracy, and F1 score. Mean precision includes results with an intersection-over-union (IU) threshold of 0.5 and combined results with IU thresholds ranging from 0.5 to 0.95. For low-contrast and small-sized lesions, small-target recall, false positive rate, and localization error can be separately calculated. All comparative experiments used the same data partitioning, decision thresholds, and statistical methods.

[0141] In summary, this invention takes the spectral prototype of pneumonia lesions as its core, and organically combines amplitude-phase coupled frequency domain enhancement, bidirectional gating space-frequency mutual verification fusion of candidate lesions, and dynamic detection convolution of spectral prototype modulation. It proposes a special technical solution to address the difficulties in detecting low-contrast, blurred-boundary, diverse-morphology, and small lesions in pneumonia pathological images.

[0142] Example 2:

[0143] This embodiment provides a pneumonia pathological image detection system based on frequency domain analysis and dynamic convolution, including: a lung image acquisition and lung field constraint preprocessing unit, a suspected lesion spectrum prototype establishment unit, a lesion spectrum prototype guided frequency domain enhancement unit, a candidate suspected lesion guided bidirectional gating space-frequency fusion unit, a spectrum prototype modulated pathological perception dynamic convolution unit, a decoupled detection output unit, and a joint training unit.

[0144] In this embodiment, the lung image acquisition and lung field constraint preprocessing unit is used to acquire pneumonia pathological images, preprocess the pneumonia pathological images, and determine the lung field constraint area.

[0145] In this embodiment, the suspected lesion spectrum prototype establishment unit is used to establish a learnable lesion spectrum prototype based on the lung field constraint region.

[0146] In this embodiment, the lesion spectrum prototype-guided frequency domain enhancement unit is used to input the preprocessed pneumonia pathological image and lesion spectrum prototype into the frequency domain-spatial domain dual-branch encoder for processing, to obtain spatial features and the frequency domain enhancement feature map, spectrum anomaly map and spectrum prototype response vector output by the lesion spectrum prototype-guided frequency domain enhancement module.

[0147] In this embodiment, the candidate suspected lesion-guided bidirectional gated space-frequency fusion unit is used to perform feature fusion on the spatial domain features and the frequency domain enhanced feature map using the bidirectional gated space-frequency fusion module and the spectral anomaly map to obtain mutually verifiable candidate features;

[0148] In this embodiment, the spectrum prototype modulation pathological perception dynamic convolution unit is used to input the mutual evidence candidate features and the spectrum prototype response vector into the spectrum prototype modulation pathological perception dynamic convolution for processing. The dynamic convolution response is obtained by using the generated dynamic detection convolution kernel, and the candidate detection feature is output after adding the dynamic convolution response to the mutual evidence candidate feature residual.

[0149] In this embodiment, the decoupling detection output unit is used to process candidate detection features using the decoupling detection head and output the final pneumonia lesion detection result;

[0150] In this embodiment, the joint training unit is used to perform joint training using detection loss, spectral prototype constraint loss, phase fidelity loss, spatial-frequency mutual verification consistency loss, and dynamic kernel smoothing loss.

[0151] The aforementioned system can implement the pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution as described in Embodiment 1. However, the implementation device of the pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution as described in Embodiment 1 includes, but is not limited to, the pneumonia pathological image detection system based on frequency domain analysis and dynamic convolution as described in this invention.

[0152] Example 3:

[0153] This invention provides an electronic device, including a processor and a memory. The memory stores program instructions for implementing the pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution of Embodiment 1. The processor executes the program instructions stored in the memory to perform pneumonia pathological image detection.

[0154] The processor can also be called a Central Processing Unit (CPU). A processor may be an integrated circuit chip with signal processing capabilities. A processor can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0155] The methods described in Embodiment 1 of this invention are applicable to this electronic device embodiment. The specific functions implemented by this electronic device embodiment are the same as those in the above method embodiment, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0156] Example 4:

[0157] This embodiment provides a storage medium storing a computer program or instructions capable of implementing the above-described pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution. When the computer program or instructions are run on a computer, the steps of the above-described pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution are executed.

[0158] The computer program or instructions can be stored in the aforementioned storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0159] The methods described in Embodiment 1 of this invention are all applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0160] This invention provides a method and system for detecting pneumonia pathological images based on frequency domain analysis and dynamic convolution. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for detecting pneumonia pathological images based on frequency domain analysis and dynamic convolution, characterized in that, Includes the following steps: Step 1: Acquire pneumonia pathological images, preprocess the pneumonia pathological images and determine the lung field constraint region, and establish a learnable lesion spectrum prototype based on the lung field constraint region; Step 2: Input the preprocessed pneumonia pathological image and lesion spectrum prototype into the frequency domain-spatial domain dual-branch encoder for processing to obtain spatial features and the frequency domain enhancement feature map, spectrum anomaly map and spectrum prototype response vector output by the frequency domain enhancement module guided by the lesion spectrum prototype; Step 3: Using the candidate lesion to guide the bidirectional gated space-frequency fusion module, the spatial domain features and the frequency domain enhanced feature map are fused together with the spectrum anomaly map to obtain mutual verification candidate features; Step 4: Input the mutual verification candidate features and the spectrum prototype response vector into the spectrum prototype modulation pathological perception dynamic convolution module for processing. Use the generated dynamic detection convolution kernel to obtain the dynamic convolution response, and add the dynamic convolution response to the mutual verification candidate features by residual addition to output the candidate detection features. Step 5: Process the candidate detection features using the decoupled detection head and output the final pneumonia lesion detection results.

2. The pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution according to claim 1, characterized in that, In step 1, the spectrum prototype includes the amplitude spectrum, phase consistency spectrum, and direction spectrum of the lesion area and the normal lung texture area.

3. The pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution according to claim 1, characterized in that, In step 2, the process of the lesion spectrum prototype-guided frequency domain enhancement module processing features includes: A two-dimensional fast Fourier transform is performed on the input feature map to obtain the complex spectrum, which is then separated into the amplitude spectrum and the phase spectrum. The spectrum region is divided by a learnable annular-sector spectral mask, and each region is matched with the lesion spectrum prototype to obtain the matching result. Based on the matching results, amplitude enhancement gating and phase fidelity gating are generated. The amplitude spectrum and phase spectrum are processed by amplitude enhancement gating and phase fidelity gating respectively to obtain the enhanced amplitude spectrum and the constrained phase spectrum. The enhanced amplitude spectrum and the constrained phase spectrum are recombined into a complex spectrum, which is then subjected to inverse Fourier transform to obtain a frequency domain enhanced feature map. At the same time, a spectral anomaly map and a spectral prototype response vector are output.

4. The pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution according to claim 1, characterized in that, Step 3 includes: The spatial features are processed to generate an initial candidate heatmap; the spectral anomaly map output by the frequency domain enhancement module guided by the lesion spectrum prototype in step 2 is reused, and a scale reliability map is generated according to the coding level of the spatial features and the size of the candidate box. The space-guided frequency-domain gating branch receives spatial features, frequency-domain enhanced feature maps, candidate heatmaps, spectral anomaly maps, and scale reliability maps to generate a space-guided frequency-domain gating. The frequency-domain enhanced feature map is weighted using this gating and added to the spatial feature residual to obtain the first branch feature. The frequency-domain-guided spatial gating branch generates a frequency-domain-guided spatial gating based on the spectral anomaly map and the spectral prototype response vector. The spatial features are weighted using this gating and added to the frequency-domain enhanced feature map residual to obtain the second branch feature. The first branch feature and the second branch feature are added together to obtain the bidirectional fusion base; the first branch feature and the second branch feature are subjected to mutual verification consistency judgment to generate consistency weight; the consistency weight is applied to the bidirectional fusion base to generate lesion enhancement term; the supplementary amount of the consistency weight is applied to the bidirectional fusion base to generate background suppression term; the lesion enhancement term is superimposed on the bidirectional fusion base and the background suppression term is subtracted to obtain mutual verification candidate features.

5. The pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution according to claim 1, characterized in that, Step 4 includes: First, candidate morphological vectors are extracted from spatial features. Spectral response vectors are extracted from the spectral prototype response vectors output by the frequency domain enhancement module guided by the lesion spectral prototype. Scale vectors are then generated based on the current coding level and candidate box size. Then, the candidate morphological vector, spectral response vector and scale vector are concatenated and input into the kernel modulation network for processing to generate basic convolutional kernel combination weights, which include block convolutional kernel weights, directional convolutional kernel weights, dilated convolutional kernel weights and high-resolution small lesion convolutional kernel weights. Next, the preset basic convolutional kernel group is adaptively weighted and summed using the weights of the basic convolutional kernel combination to generate a dynamic detection convolutional kernel; Finally, the candidate features for mutual verification are convolved using a dynamic detection convolution kernel to obtain a dynamic convolution response. The dynamic convolution response is then added to the residual of the candidate features for mutual verification to obtain the candidate detection features.

6. The pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution according to claim 1, characterized in that, Also includes: End-to-end training is performed using a joint loss function, which includes detection loss, spectral prototype constraint loss, phase fidelity loss, spatial-frequency mutual verification consistency loss, and dynamic kernel smoothing loss; the detection loss includes classification loss, bounding box regression loss, and confidence loss.

7. A pneumonia pathological image detection system based on frequency domain analysis and dynamic convolution, characterized in that, include: The lung image acquisition and lung field constraint preprocessing unit is used to acquire pneumonia pathological images, preprocess the pneumonia pathological images, and determine the lung field constraint area. The suspected lesion spectrum prototype establishment unit is used to establish a learnable lesion spectrum prototype based on the lung field constraint region. The lesion spectrum prototype-guided frequency domain enhancement unit is used to input the preprocessed pneumonia pathological image and lesion spectrum prototype into the frequency domain-spatial domain dual-branch encoder for processing, to obtain spatial features and the frequency domain enhancement feature map, spectrum anomaly map and spectrum prototype response vector output by the lesion spectrum prototype-guided frequency domain enhancement module; The candidate suspected lesion guides a bidirectional gated space-frequency fusion unit, which is used to fuse spatial domain features and frequency domain enhanced feature maps using a bidirectional gated space-frequency fusion module and a spectral anomaly map to obtain mutually verifiable candidate features; The Spectrum Prototype Modulation Pathological Perception Dynamic Convolution Unit is used to input the mutual evidence candidate features and the spectral prototype response vector into the Spectrum Prototype Modulation Pathological Perception Dynamic Convolution for processing. The generated dynamic detection convolution kernel is used to obtain the dynamic convolution response, and the dynamic convolution response is added to the mutual evidence candidate feature residual to output the candidate detection feature. The decoupled detection output unit is used to process candidate detection features using the decoupled detection head and output the final pneumonia lesion detection result; The joint training unit is used for joint training using detection loss, spectral prototype constraint loss, phase fidelity loss, spatial-frequency mutual verification consistency loss, and dynamic kernel smoothing loss.

8. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The device stores a computer program or instructions that, when executed on a computer, perform the steps of the pneumonia pathological image detection method based on frequency domain analysis and dynamic convolution as described in any one of claims 1 to 6.

Images