Difficult airway assessment system and device using MRI images in combination with deep learning algorithms

Abstract

This invention belongs to the field of novel medical device technology at the intersection of medical and artificial intelligence equipment. It relates to a difficult airway assessment system and device that utilizes MRI images combined with deep learning algorithms. The system includes: a heatmap generation module for acquiring the original two-dimensional MRI image slice sequence of the subject to be assessed and generating a saliency heatmap corresponding to each original slice; a mask image segmentation module for generating a soft probability ROI mask sequence based on each saliency heatmap and the original two-dimensional MRI image slice sequence; and a difficult airway prediction and assessment module for outputting the difficult airway prediction and assessment result of the subject based on the original two-dimensional MRI image slice sequence and the corresponding soft probability ROI mask sequence. This invention effectively overcomes problems such as difficulties in soft tissue quantification, insufficient segmentation accuracy, and manual measurement bias, achieving deep integration of global structure and local details, thus improving the effectiveness and accuracy of difficult airway prediction and assessment.

Description

Technical Field

[0001] This invention relates to the field of novel medical device technology that combines medical and artificial intelligence equipment, and in particular to a difficult airway assessment system and device that utilizes MRI images combined with deep learning algorithms. Background Technology

[0002] Endotracheal intubation is a key technique in anesthesiology, emergency medicine, and critical care medicine. Preoperative airway assessment is crucial for screening patients with difficult airways and reducing ventilation difficulties and the risk of asphyxiation. Currently, assessment of difficult airways mainly relies on physical examination, imaging techniques, and ultrasound technology. Morphological measurements of bony and soft tissue structures are used to comprehensively determine the difficulty of intubation.

[0003] In recent years, the combination of artificial intelligence and radiology has significantly improved prediction accuracy. However, existing solutions still have the following shortcomings: First, most methods only analyze a single image or a single anatomical structure, ignoring the multi-layered and multi-regional overall features of the soft tissues of the head and neck, especially lacking accurate assessment methods for areas with blurred soft tissue edges and large morphological variations; Second, geometric parameters based on manual annotation are easily affected by the operator's subjectivity and are difficult to fully capture the spatial distribution and morphological details of soft tissues; Third, there is currently no difficult airway prediction model that can deeply integrate global anatomy and local fine morphological features, resulting in poor prediction performance and accuracy. Summary of the Invention

[0004] Based on the above analysis, the present invention aims to provide a difficult airway assessment system and device that utilizes MRI images combined with deep learning algorithms to solve the problems of existing technologies that rely on manual labor, ignore soft tissue features, lack global and local feature fusion mechanisms, and have poor assessment quality.

[0005] On one hand, embodiments of the present invention provide a difficult airway assessment system utilizing MRI images combined with deep learning algorithms, comprising: A heatmap generation module is used to acquire the original two-dimensional MRI image slice sequence of the subject to be evaluated and generate a saliency heatmap corresponding to each original slice; wherein the saliency heatmap is used to indicate at least the key soft tissue areas of the upper airway, the base of the tongue and the epiglottis. The mask image segmentation module is used to generate a soft probability ROI mask sequence based on each of the saliency heatmaps and the original two-dimensional MRI image slice sequence; wherein each mask in the soft probability ROI mask sequence is a multi-channel probability map, and different channel probability maps correspond to the confidence probability of the upper airway, tongue base and epiglottic region respectively. The difficult airway prediction and assessment module is used to output the difficult airway prediction and assessment results of the subject under assessment based on the original two-dimensional MRI image slice sequence and the corresponding soft probability ROI mask sequence.

[0006] Furthermore, the heatmap generation module is implemented based on the CAM-ResNet-18 generation model, which integrates the CAM class activation mapping algorithm and the ResNet-18 residual network; wherein the CAM-ResNet-18 generation model includes at least: shallow residual blocks for outputting high spatial resolution, low channel number feature maps, and deep residual blocks for outputting low spatial resolution, high channel number feature maps.

[0007] Furthermore, the heatmap generation module includes: a saliency heatmap generation unit; the saliency heatmap generation unit is used for: Based on the shallow residual blocks and deep residual blocks, feature extraction is performed on the original two-dimensional MRI image slice sequence, and multi-dimensional feature maps with different spatial resolutions and channel numbers corresponding to a single original slice are obtained. Based on the CAM-type activation mapping algorithm, multiple initial heatmaps of different resolutions are generated for each original slice's corresponding multidimensional feature map. Each initial heatmap corresponding to each original slice is upsampled and unified to the original resolution corresponding to each original slice to obtain multiple upsampled heatmaps. The multiple upsampled heatmaps are weighted and summed according to preset fusion weights to obtain the saliency heatmap corresponding to each original slice.

[0008] Furthermore, the heatmap generation module also includes: an evaluation and verification unit for determining whether the saliency heatmap satisfies the dual verification path; The dual verification path includes: Qualitative verification is used to overlay each of the aforementioned saliency heatmaps with their corresponding original slices in a semi-transparent pseudo-color manner, and based on visual examination information input by clinical experts, to determine whether the key soft tissue areas in the overlaid images match the actual anatomical locations of the tongue base, epiglottis, and upper airway. Quantitative evaluation is used to compare the significance heatmap corresponding to each original slice with the discrete binary mask manually annotated by clinical experts to determine whether the comparison result meets the preset threshold condition. The evaluation and verification unit is further configured to transmit the saliency heatmap obtained through dual evaluation to the mask image segmentation module after both the qualitative verification and the quantitative evaluation have passed the evaluation.

[0009] Furthermore, the mask image segmentation module is implemented based on an enhanced CAM-U-Net segmentation model that integrates the CAM class activation mapping algorithm and the U-Net network; wherein, the CAM-U-Net segmentation model includes at least: a U-Net encoder for extracting multi-scale feature maps, a U-Net decoder with different spatial resolution levels, and a spatial attention residual module for performing spatial attention enhancement processing.

[0010] Furthermore, the mask image segmentation module includes at least: a heatmap prior pyramid unit and a feature stitching unit; The heatmap prior pyramid unit is used to: use the saliency heatmap corresponding to each original slice in the original two-dimensional MRI image slice sequence as the bottom layer of the pyramid; use the bilinear interpolation algorithm to dynamically downsample each saliency heatmap to generate sub-heatmaps that match the feature map size of each level of the U-Net decoder, so as to form the heatmap prior pyramid; The feature stitching unit is used to: stitch the feature map of the U-Net decoder after upsampling at the previous level with the skip connection feature map of the corresponding level of the U-Net encoder along the channel dimension to obtain the original decoding features of each layer; introduce a matching sub-heatmap at each level of the U-Net decoder, and input each sub-heatmap into the spatial attention residual module to fuse it with the original decoding features of the corresponding level to obtain a multi-channel stitched feature map.

[0011] Furthermore, the spatial attention residual module is also used to perform the following operations on each level of the U-Net decoder: Receive the multi-channel stitched feature map corresponding to the current level and the sub-heatmap that matches the spatial resolution of the level; The sub-heatmap is input into a convolutional layer and processed by a Sigmoid activation function to generate a spatial attention weight map. The spatial attention weight map is broadcast by channel and then multiplied element-wise with the multi-channel stitched feature map to obtain the enhanced feature map; The enhanced feature map is added element-wise to the multi-channel stitched feature map to output the final feature map of this level. The final feature map is then convolved to output a soft probability ROI mask. Based on the soft probability ROI mask corresponding to each original slice, the soft probability ROI mask sequence is obtained.

[0012] Furthermore, the difficult airway prediction and evaluation module is implemented based on a three-dimensional SE-ResNet-18 prediction model that integrates the SE compression-excitation channel attention mechanism with the ResNet-18 residual network. The three-dimensional SE-ResNet-18 prediction model includes at least: a first branch for extracting global anatomical features, a second branch for extracting local anatomical features, a gated attention fusion module for adaptively fusing the output features of the two branches, and a Softmax classification layer.

[0013] Furthermore, the difficult airway prediction and assessment module is used for: Based on the first branch, the three-dimensional voxel matrix obtained by re-stacking the original two-dimensional MRI image slice sequence of the object to be evaluated along the physical space order is used as input to extract global anatomical features and output a one-dimensional global feature vector. Based on the second branch, the three-dimensional soft probability mask volume obtained by re-stacking the soft probability ROI mask sequence of the object to be evaluated along the physical space order is used as input to extract local morphological features and output a one-dimensional local feature vector. Based on the gated attention fusion module, the one-dimensional global feature vector and the one-dimensional local feature vector are concatenated along the channel dimension to obtain a concatenated feature vector; a gating coefficient is generated based on the concatenated feature vector, and the one-dimensional global feature vector and the one-dimensional local feature vector are weighted and summed using the gating coefficient to obtain a fused feature vector; the fused feature vector is input to the Softmax classification layer to output the classification prediction probability of the difficult airway, so as to obtain the prediction evaluation result of the difficult airway.

[0014] On the other hand, embodiments of the present invention provide a computer device, including at least one processor and at least one memory communicatively connected to the processor; The memory stores instructions executable by the processor to implement a difficult airway assessment system using MRI images combined with deep learning algorithms as described in any of the preceding claims.

[0015] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: First, unlike related technologies that rely on single-parameter measurement and are greatly affected by human subjectivity, this invention is based on the head and neck MRI images of the subject to be evaluated. It can automatically identify areas with blurred soft tissue edges and large morphological variations using saliency heatmaps, and output soft probability masks corresponding to key soft tissue areas such as the upper airway, tongue base and epiglottis. This effectively overcomes problems such as difficulty in soft tissue quantification, insufficient segmentation accuracy and manual measurement deviation.

[0016] Second, unlike related technologies which lack a mechanism for fusing global anatomical and local morphological features, this invention uses a weight-sharing dual-branch 3D network to extract global anatomical features (such as the overall topology of the cervical spine and mandible) and local geometric features (such as the narrowest cross section of the airway and the volume of the tongue root). By adaptively adjusting the contribution ratio of the two types of features through a gated attention fusion module, it achieves deep fusion of global structure and local details, avoids the limitations of a single feature source, and significantly improves the effect and accuracy of predicting difficult airways. Furthermore, this invention overcomes the limitations of traditional manual operation or semi-automated execution, utilizing intelligent devices for fully automated implementation without relying on human intervention. This provides a new model for assisted medical care, achieving a fully automated closed loop from data analysis to evaluation result generation. The output results are used to assist clinical medical care, and the final diagnosis and treatment are confirmed and executed by the physician. This enhances the system's information processing capabilities, thereby improving overall evaluation efficiency and quality.

[0017] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0018] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a schematic diagram of the main modules of the difficult airway assessment system utilizing MRI images combined with deep learning algorithms according to an embodiment of the present invention; Figure 2 This is a schematic diagram of CAM class activation mapping and ResNet-18 network according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the CAM-U-Net segmentation model according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the three-dimensional SE-ResNet-18 prediction model according to an embodiment of the present invention. Detailed Implementation

[0019] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0020] One specific embodiment of the present invention proposes a difficult airway assessment system utilizing MRI images combined with deep learning algorithms, such as... Figure 1 As shown, it specifically includes the following modules: A heatmap generation module is used to acquire the original two-dimensional MRI image slice sequence of the subject to be evaluated and generate a saliency heatmap corresponding to each original slice; wherein the saliency heatmap is used to indicate at least the key soft tissue areas of the upper airway, the base of the tongue and the epiglottis. The mask image segmentation module is used to generate a soft probability ROI mask sequence based on each of the saliency heatmaps and the original two-dimensional MRI image slice sequence; wherein each mask in the soft probability ROI mask sequence is a multi-channel probability map, and different channel probability maps correspond to the confidence probability of the upper airway, tongue base and epiglottic region respectively. The difficult airway prediction and assessment module is used to output the difficult airway prediction and assessment results of the subject under assessment based on the original two-dimensional MRI image slice sequence and the corresponding soft probability ROI mask sequence.

[0021] To achieve rapid, accurate, and intelligent identification of difficult airways during implementation, this invention designs an end-to-end deep learning framework encompassing image anatomical localization, fine pixel segmentation, and intelligent risk prediction. First, original two-dimensional MRI image slice sequences of the subject to be evaluated are acquired using an MRI scanner. Based on these MRI images (Magnetic Resonance Imaging), a saliency heatmap indicating key soft tissue regions is generated using the CAM algorithm (Class Activation Mapping). Second, the original image of the subject to be evaluated and the saliency heatmap are used together as input to the segmentation module to obtain a soft probability ROI mask. Finally, the soft probability ROI mask (Region of Interest) and the original image are used together as input to the prediction and evaluation model, ultimately outputting the difficult airway prediction and evaluation result. The aforementioned original two-dimensional MRI image is a head and neck MRI image of the subject to be evaluated, which includes, but is not limited to, the soft tissue anatomical structures of the upper airway region, the base of the tongue region, and the epiglottis region.

[0022] It is understood that a difficult airway refers to a situation in which a physician who has received clinical anesthesia training encounters a foreseeable or unforeseen difficulty or failure in airway management in clinical practice; therefore, the difficult airway prediction and assessment results of this embodiment of the invention refer to an assessment index that can characterize the risk level or probability of the subject being assessed developing a difficult airway.

[0023] Meanwhile, the difficult airway prediction and assessment results of this embodiment of the invention can be binary probability values, for example, when the probability is greater than or equal to a judgment threshold, it is determined to be a difficult airway; otherwise, it is a non-difficult airway. Alternatively, it can be multiple risk levels, for example, when using a multi-class output head, it can output "low risk," "medium risk," and "high risk" classifications. It can also be continuous scoring, for example, outputting a score from 0 to 10, with higher scores representing higher airway difficulty. Of course, the specific form of the difficult airway prediction and assessment results depends on the design of the output layer, and is not limited here.

[0024] It should be noted that this invention does not involve professional disease diagnosis and treatment, but only provides a data information processing and automated assessment method. The output results of this system are all auxiliary references, and the final decision is made by the doctor based on clinical standards.

[0025] The following section provides a detailed description of the components and functions of each module in the difficult airway assessment system utilizing MRI images combined with deep learning algorithms as described in this invention.

[0026] First, the heatmap generation module.

[0027] In some implementations, such as Figure 2 As shown, the heatmap generation module is implemented based on the CAM-ResNet-18 generation model, which integrates the CAM class activation mapping algorithm and the ResNet-18 residual network. The CAM-ResNet-18 generation model includes at least: shallow residual blocks for outputting high spatial resolution and low channel number feature maps, and deep residual blocks for outputting low spatial resolution and high channel number feature maps.

[0028] In practice, direct segmentation of soft tissues (such as the upper airway, tongue base, and epiglottis) in head and neck MRI images is challenging due to their blurred edges. Therefore, a pre-trained ResNet-18 deep neural network model can be selected as the basic feature extraction network, with the CAM (Class Activation Mapping) algorithm serving as the visual interpreter. The synergy between the two lies in the fact that the ResNet-18 network extracts hidden multidimensional features from the image, while the CAM algorithm transforms these features into intuitive, non-manually accessible anatomical spatial location information, such as CAM heatmaps.

[0029] During the model training phase, the input samples of the CAM-ResNet-18 generative model include a sequence of historical two-dimensional MRI image slices with manually labeled difficult airway categories; the output samples include historical saliency heatmaps corresponding to the input slices (generated by a gradient-weighted class activation mapping algorithm), which will not be elaborated here.

[0030] In some preferred embodiments, the heatmap generation module includes: a saliency heatmap generation unit; the saliency heatmap generation unit is used for: Based on the shallow residual blocks and deep residual blocks, feature extraction is performed on the original two-dimensional MRI image slice sequence, and multi-dimensional feature maps with different spatial resolutions and channel numbers corresponding to a single original slice are obtained. Based on the CAM-type activation mapping algorithm, multiple initial heatmaps of different resolutions are generated for each original slice's corresponding multidimensional feature map. Each initial heatmap corresponding to each original slice is upsampled and unified to the original resolution corresponding to each original slice to obtain multiple upsampled heatmaps. The multiple upsampled heatmaps are weighted and summed according to preset fusion weights to obtain the saliency heatmap corresponding to each original slice.

[0031] Taking Grad-CAM (Gradient-Weighted Class Activation Mapping) as an example, its specific operation process is as follows: First, slice a single two-dimensional MRI image (size denoted as...). ,like The input (pixel) is a pre-trained 2D ResNet-18 residual network. The network's four residual blocks, such as layer 1 to layer 4, sequentially output multidimensional feature maps, each containing three dimensions: spatial height dimension... Spatial width dimension and channel dimensions .

[0032] Among them, the shallow residual blocks (layer1, layer2) output feature maps with high spatial resolution and low channel count (e.g., layer1 outputs...). layer2 output ), responsible for capturing basic edges and textures; deep residual blocks (layer3, layer4) output low spatial resolution, high channel number feature maps (e.g., layer3 output). Layer 4 output ), responsible for capturing high-dimensional semantics.

[0033] Subsequently, based on the multidimensional feature maps output by each residual block, the gradient of the target category (i.e., the "difficult airway" category in the target classification task, such as difficult and non-difficult) with respect to each feature map channel is calculated using the Grad-CAM algorithm. The weights of each channel are obtained through global average pooling, followed by spatial weighted summation and ReLU activation, outputting the initial resolution saliency heatmaps for the corresponding shallow (layer1, layer2) and deep (layer3, layer4) layers. The original resolution of each layer's heatmap is consistent with the spatial resolution of the corresponding residual block's output feature map. Furthermore, the shallow and deep heatmaps are output independently without forced fusion; subsequent multi-scale fusion or separate use can be selected as needed.

[0034] Finally, each low-resolution heatmap is uniformly enlarged to the size of the original input slice by bilinear interpolation upsampling. ,Pick (pixels). The four upsampled heatmaps mentioned above are not all directly output. Instead, they are weighted and summed using the fusion weights determined during the hyperparameter optimization stage to obtain a final saliency heatmap (size 256×256 pixels, values ​​[0,1] with continuous probability). This final heatmap is the only output and is used as input to the subsequent mask image segmentation module. Preferably, the fusion weights are four positive numbers that sum to 1, and they are determined through Bayesian search during the optimization stage.

[0035] Furthermore, to verify the positioning accuracy of the saliency heatmap generated by CAM, this stage adopts two independent verification paths: qualitative visualization verification and quantitative index evaluation.

[0036] Preferably, the heatmap generation module further includes: an evaluation and verification unit for determining whether the saliency heatmap satisfies the dual verification path; Specifically, the dual verification path includes: Qualitative verification involves overlaying each of the aforementioned saliency heatmaps in a semi-transparent pseudo-color manner onto its corresponding original slice. Based on visual inspection information input by clinical experts, it determines whether the key soft tissue regions in the overlaid image correspond to the actual anatomical locations of the tongue base, epiglottis, and upper airway. For example, the upsampled saliency heatmaps are overlaid in a semi-transparent pseudo-color manner onto the same original MRI image slice to generate an overlaid image. At least two clinical experts independently conduct visual inspections, such as checking whether the highlighted areas of the heatmaps basically correspond to the anatomical locations of the tongue base, epiglottis, and upper airway, and whether there are any significant deviations. Visual inspection information is then input.

[0037] Quantitative evaluation is used to compare the saliency heatmap corresponding to each original slice with a discrete binary mask manually annotated by clinical experts to determine whether the comparison result meets the preset threshold condition; for example, the saliency heatmap of continuous probability after upsampling is compared with the discrete binary mask manually annotated by clinical experts (as an artificial prior standard for evaluation reference).

[0038] Preferably, during manual annotation, experienced clinical experts strictly adhere to unified anatomical definitions, using polygon or pen tools to manually and precisely delineate the intraluminal boundaries of the upper airway, the outer contour of the tongue base, and the cartilaginous structure of the epiglottis on each MRI image slice, obtaining a corresponding multi-category indexed mask image. Pixel values ​​of 0 represent the background, 1 the upper airway, 2 the tongue base, and 3 the epiglottis. Since the CAM heatmap at this stage is generated for the image-level classification category of "difficult airway," its goal is to highlight all anatomical regions related to the difficult airway (i.e., the union of the upper airway, tongue base, and epiglottis) in the entire image, rather than distinguishing specific structures. Therefore, for quantitative evaluation, the above multi-category indexed mask is first converted into a binary mask: any pixel value greater than 0 (i.e., belonging to any of the upper airway, tongue base, or epiglottis categories) is... Otherwise, it is 0. This binary mask represents the overall range of the key anatomical region confirmed by the expert. In addition, since the heatmap is a continuous probability and the expert mask is a binary image, the heatmap can be binarized with a threshold of 0.5 to obtain the prediction mask, and then its IoU (Intersection over Union) with the expert mask can be calculated. The preset acceptance threshold is IoU ≥ 0.7.

[0039] Furthermore, the evaluation and verification unit is also used to transmit the saliency heatmap, which has passed both the qualitative verification and the quantitative evaluation, to the mask image segmentation module after both verifications have passed. That is, only when both verification paths pass can the final saliency heatmap output by the model be input into the subsequent segmentation module for inference.

[0040] It should be noted that the evaluation and verification unit is mainly used in the model training stage. It can optimize the relevant model parameters based on the historical significance heatmap to obtain a better parameter configuration. In the actual application stage of the heatmap generation module, the evaluation and verification unit is not a necessary component, but it can be retained according to the specific scenario requirements and continue to perform heatmap evaluation and verification functions. There are no restrictions here.

[0041] Second, the mask image segmentation module.

[0042] After obtaining the CAM saliency heatmap with accurate localization in the previous stage, this stage uses it as spatial prior knowledge and injects it into the U-Net segmentation network. In this way, the pre-constructed enhanced CAM-U-Net segmentation model solves the problem that the boundaries of soft tissues such as airways and tongue base in head and neck MRI images are extremely blurred and difficult to segment accurately by ordinary networks.

[0043] In some implementations, such as Figure 3 As shown, the mask image segmentation module is implemented based on an enhanced CAM-U-Net segmentation model that integrates the CAM class activation mapping algorithm and the U-Net network; wherein, the CAM-U-Net segmentation model includes at least: a U-Net encoder for extracting multi-scale feature maps, a U-Net decoder with different spatial resolution levels, and a spatial attention residual module for performing spatial attention enhancement processing.

[0044] During the training phase, the input samples of the CAM-U-Net segmentation model include historical two-dimensional MRI image slice sequences and historical saliency heatmaps pre-calculated by the CAM-ResNet-18 generation model. Manually labeled pixel-level soft probability ROI mask sequences are used as ground values ​​for supervised learning, which will not be elaborated here.

[0045] In some preferred embodiments, the mask image segmentation module includes at least: a heatmap prior pyramid unit and a feature stitching unit; The heatmap prior pyramid unit is used to: use the saliency heatmap corresponding to each original slice in the original two-dimensional MRI image slice sequence as the bottom layer of the pyramid; use the bilinear interpolation algorithm to dynamically downsample each saliency heatmap to generate sub-heatmaps that match the feature map size of each level of the U-Net decoder, so as to form the heatmap prior pyramid; The feature stitching unit is used to: stitch the feature map of the U-Net decoder after upsampling at the previous level with the skip connection feature map of the corresponding level of the U-Net encoder along the channel dimension to obtain the original decoding features of each layer; introduce a matching sub-heatmap at each level of the U-Net decoder, and input each sub-heatmap into the spatial attention residual module to fuse it with the original decoding features of the corresponding level to obtain a multi-channel stitched feature map.

[0046] Specifically, during the execution of the heatmap prior pyramid unit, the CAM saliency heatmap output from the previous stage is a single-channel image with the exact same slice size as the original input MRI image, and its size is... ,Pick Pixel; each pixel has a continuous probability of [0,1], representing the significance of the location belonging to the difficult airway-related anatomical region.

[0047] Continue to refer to Figure 3 As shown, since the standard U-Net decoder (upsampling stage) contains multiple layers with different spatial resolutions, the corresponding feature map sizes are as follows: , , , Pixels, etc. To ensure that the aforementioned single-channel CAM heatmap can play a guiding role at each decoder level, the method for constructing the heatmap prior pyramid in this embodiment of the invention includes: First, using the original high-resolution single-channel CAM saliency heatmap (size, etc.) output from the previous stage... The base of the pyramid is (pixels); secondly, the heatmap is dynamically downsampled using bilinear interpolation to generate sub-heatmaps that perfectly match the feature map size of each decoder layer of U-Net, for example, generating a size of (pixels). , , A downsampled heatmap of pixels. This results in a multi-scale set of heatmaps, which serves as the heatmap prior pyramid.

[0048] Furthermore, unlike the U-Net network which relies solely on image features extracted layer by layer by the encoder for decoding, in this embodiment, for each layer of the U-Net decoder, the upsampled feature map of the previous layer is first concatenated with the skip connection feature map of the corresponding layer of the U-Net encoder to form the original decoding features of the current layer; then, a CAM sub-heatmap with the same spatial resolution as that layer is extracted from the heatmap prior pyramid and input into the spatial attention residual module to be fused and enhanced with the original decoding features. This guides the network to focus on the highlighted areas of the heatmap (i.e., the union of the upper airway, tongue root, and epiglottis), thereby effectively overcoming the segmentation difficulties caused by the blurring of soft tissue edges.

[0049] Furthermore, the spatial attention residual module is used to perform the following operations on each level of the U-Net decoder: Receive the multi-channel stitched feature map corresponding to the current level and the sub-heatmap that matches the spatial resolution of the level; The sub-heatmap is input into a convolutional layer and processed by a Sigmoid activation function to generate a spatial attention weight map. The spatial attention weight map is broadcast by channel and then multiplied element-wise with the multi-channel stitched feature map to obtain the enhanced feature map; The enhanced feature map is added element-wise to the multi-channel stitched feature map to output the final feature map of this level. The final feature map is then convolved to output a soft probability ROI mask. Based on the soft probability ROI mask corresponding to each original slice, the soft probability ROI mask sequence is obtained.

[0050] Specifically, at each layer of the CAM-U-Net decoder, it not only processes the image feature maps (multi-channel stitched feature maps, denoted as...) from the previous layer... , shape ,in For the number of channels, (For the current layer spatial resolution), it also needs to receive the CAM sub-heatmap of the corresponding size (single channel, denoted as ). , shape That is, the feature map here. It is a feature map inside the U-Net decoder, which is the result of splicing the upsampled output of the previous decoding layer and the corresponding skip connection of the encoder.

[0051] To fuse the multi-channel stitched feature map and the sub-heatmap, the specific computation process of the spatial attention residual module in this embodiment of the invention is as follows: (1) Generate attention weights: generate single-channel CAM sub-heatmaps Enter one A linear mapping is performed in the convolutional layer, followed by a Sigmoid activation function to obtain the spatial attention weight map. This convolutional layer does not change the spatial dimensions, only mapping the number of channels from 1 to 1 (i.e., maintaining a single channel), but it can introduce learnable linear transformation parameters, allowing the network to adaptively adjust the distribution of heatmap values. Simultaneously, the Sigmoid function can strictly compress its output to the range [0,1], generating a two-dimensional spatial attention weight map. (The shape remains the same) In the image, the pixel values ​​near the target airway, tongue root, and epiglottis are close to 1 (highlight area), while the pixel values ​​in the irrelevant background area are close to 0 (dark area).

[0052] (2) Soft gating feature enhancement: Using element-wise matrix multiplication, the generated two-dimensional attention weight map is enhanced. The feature map after channel broadcasting is stitched together with the multi-channel feature map of the current layer of the U-Net network. Multiplication. Specifically, the broadcast operation will multiply the shape into... of Copy along channel dimension The portion is used to obtain a shape of broadcast post-tensor This ensures that the feature map on each channel is multiplied by the same spatial attention weight map.

[0053] For example, if the current layer feature map There are 64 channels ( ), then after broadcast All 64 channels have the same spatial weight distribution. (Element-wise multiplication) In the result, the features of the target region are preserved or amplified (multiplied by a number close to 1), while the features of background noise and irrelevant tissues (such as the brain and external air) are forcibly suppressed or reduced to zero (multiplied by a number close to 0), thereby achieving a soft gating effect that enhances the features of interest and weakens the irrelevant features.

[0054] (3) Residual Feature Compensation: To avoid the problem of accidental deletion of edge image features due to slight deviations in CAM heatmap positioning, a residual connection is introduced after the multiplication operation. Specifically, the enhanced feature map... With the original feature map that has not been multiplied (That is, the multi-channel stitched feature map input to this layer, which has not undergone attention modulation) is added element-wise to obtain the module output. This residual connection allows the network to absorb key information from the prior heatmap while retaining complete original contextual image information. Even if there are slight shifts in the CAM heatmap, they can still be compensated for by the original feature map.

[0055] With input image size as Taking pixels as an example, the CAM-U-Net decoder contains four layers, with resolutions of [resolution values ​​to be filled in]. , , , Pixels. The operation sequence for each level is as follows: First, the feature map output after upsampling from the previous layer is received; second, this feature map is concatenated along the channel with the skip connection feature map of the corresponding layer of the encoder to obtain the image feature map of this layer. Then, a CAM sub-heatmap matching the resolution of that level is extracted from the heatmap prior pyramid. The enhanced feature map is obtained through the spatial attention residual module. Finally, Enter two The convolutional layer performs further feature extraction and upsamples the data to the next layer's resolution using bilinear interpolation (e.g., from...). Pixel upsampling (pixels), as one of the inputs to the next layer; repeat the above process until the highest resolution layer is... Each pixel completes the decoder's progressive feature enhancement across all levels.

[0056] Through the aforementioned layer-by-layer decoding and upsampling method with spatial prior supervision, the CAM-U-Net segmentation model outputs a high-precision soft probabilistic ROI mask in the final layer using a Softmax classifier; this mask is a three-channel probability map (size...). In this embodiment For example, each channel corresponds to an anatomical structure: the first channel is the upper airway, the second channel is the base of the tongue, and the third channel is the epiglottis. The value of each pixel in the mask is a continuous floating-point number between [0,1], representing the confidence level that the pixel belongs to the corresponding structure.

[0057] Of course, if multiple key soft tissue regions are included, the number of channels in the soft probability ROI mask is the same as the number of key soft tissue regions, and each channel corresponds to the probability distribution of a soft tissue region, which will not be elaborated here.

[0058] Therefore, unlike traditional manually drawn discrete hard masks with values ​​of 0 or 1, the soft probabilistic mask output by this invention can finely characterize the edge transition information of soft tissue. That is, at the anatomical boundary, the pixel probability value smoothly transitions from high to low, rather than abruptly. This continuous probability expression preserves purer local anatomical morphological features, and therefore can be directly used as input for the next stage of the dual-branch three-dimensional prediction model for subsequent difficult airway classification or prediction of related clinical assessment indicators.

[0059] Third, the difficult airway prediction and assessment module.

[0060] Unlike conventional single-path classification networks or existing technologies that simply stack multimodal data, the core improvement in this stage is that a dual-branch three-dimensional architecture with weight sharing and adaptive trade-off mechanisms is designed, thereby mapping global anatomy and local key morphological features to a unified high-dimensional semantic space.

[0061] In some implementations, such as Figure 4 As shown, the difficult airway prediction and evaluation module is implemented based on a three-dimensional SE-ResNet-18 prediction model that integrates the SE compression-excitation channel attention mechanism with the ResNet-18 residual network. The three-dimensional SE-ResNet-18 prediction model includes at least: a first branch for extracting global anatomical features, a second branch for extracting local anatomical features, a gated attention fusion module for adaptively fusing the output features of the two branches, and a Softmax classification layer.

[0062] The dual-branch 3D prediction model (denoted as 3D SE-ResNet-18) in this stage is independent of the aforementioned CAM-U-Net segmentation model. The two are sequentially linked: the CAM-U-Net segmentation model is responsible for extracting a fine multi-class soft probabilistic ROI mask from the original slices, while the dual-branch 3D prediction model in this stage uses this mask as the input of the local flow and combines it with the 3D volume data after the original MRI image slices are re-stacked to complete the final classification prediction of difficult airways. For example, it outputs the binary probability (value range [0,1]) of whether a single object to be evaluated belongs to "difficult airway".

[0063] It is understandable that the prediction model contains two parallel branches with completely identical structures. The basic architecture of each branch is a 3D SE-ResNet-18 model, specifically including: a 3×3×3 three-dimensional convolutional layer, four three-dimensional bottleneck residual stages, and a three-dimensional global average pooling layer at the end; wherein, each three-dimensional bottleneck residual stage is composed of multiple stacked three-dimensional residual blocks, and each three-dimensional residual block contains at least multiple three-dimensional convolutional layers and batch normalization layers; and an SE (Squeeze-and-Excitation) compression-excitation channel attention mechanism is embedded within each three-dimensional residual block.

[0064] In some preferred embodiments, the difficult airway prediction and assessment module is used for: Based on the first branch, the three-dimensional voxel matrix obtained by re-stacking the original two-dimensional MRI image slice sequence of the object to be evaluated along the physical space order is used as input to extract global anatomical features and output a one-dimensional global feature vector. Based on the second branch, the three-dimensional soft probability mask volume obtained by re-stacking the soft probability ROI mask sequence of the object to be evaluated along the physical space order is used as input to extract local morphological features and output a one-dimensional local feature vector. Specifically, in the previous stage, the CAM-U-Net segmentation model processed each original 2D MRI image slice independently. For a single subject to be evaluated (whose head and neck MRI scan contained 12-13 consecutive 2D slices), the CAM-U-Net segmentation model output a corresponding number of 2D soft probabilistic ROI masks. Each mask is a three-channel mask (upper airway, tongue base, epiglottis) with the same size as the original slice (e.g., ...). (pixels). Therefore, a single object to be evaluated outputs a total of 12-13 two-dimensional mask sequences. In this stage, the original two-dimensional MRI image slice sequences (grayscale images) of the same object to be evaluated are re-stacked into a three-dimensional voxel matrix along the Z-axis (spatial physical order), and the soft probability ROI mask sequences output by the corresponding CAM-U-Net segmentation model are also stacked into a three-dimensional soft probability mask volume along the Z-axis.

[0065] Preferably, continue to refer to Figure 4 As shown, in the 3D SE-ResNet-18 prediction model, the dual-branch structure is used to perform the following operations: (1) Global Stream (Branch-A) Input: 12-13 original two-dimensional MRI image slices of the same object to be evaluated are re-stacked in physical spatial order (along the Z-axis) into a complete three-dimensional voxel matrix (size). ,in The number of slices. Input the first branch and extract the overall spatial topological features of the mandible, cervical spine and soft tissues through the branch.

[0066] (2) Branch-B Input: Stack 12-13 two-dimensional soft probabilistic ROI masks (each with three channels) corresponding to the same object to be evaluated, as output by the CAM-U-Net segmentation model in the previous stage, along the Z-axis to obtain a three-dimensional soft probabilistic mask volume (size). The second branch is input. This branch focuses on encoding pure geometric features such as the narrowest cross-section of the airway and the volume of the tongue root. It should be noted that before input, this mask is compressed into a single-channel 3D volume (size...) by taking the maximum value from the three channels at each spatial location. This simplifies the input dimensions and preserves the most significant structural responses.

[0067] Furthermore, the two branches mentioned above can share all the convolutional kernel parameters of the aforementioned 3D SE-ResNet-18 prediction model. This mechanism not only reduces the number of model parameters by half, but also forces the network to learn grayscale MRI images and ROI masks of purely geometric shapes, thus providing strong adaptability to both types of information.

[0068] Furthermore, the difficult airway prediction and assessment module is also used for: Based on the gated attention fusion module, the one-dimensional global feature vector and the one-dimensional local feature vector are concatenated along the channel dimension to obtain a concatenated feature vector; a gating coefficient is generated based on the concatenated feature vector, and the one-dimensional global feature vector and the one-dimensional local feature vector are weighted and summed using the gating coefficient to obtain a fused feature vector; the fused feature vector is input to the Softmax classification layer to output the classification prediction probability of the difficult airway, so as to obtain the prediction evaluation result of the difficult airway.

[0069] Specifically, the two branches output one-dimensional global feature vectors (denoted as ) through the 3D-GAP layer (three-dimensional global average pooling layer). ) and one-dimensional local eigenvectors (denoted as After that, a dedicated gated attention fusion module was designed before the final classification layer in this stage.

[0070] It is important to clarify that this gated attention fusion module is fundamentally different from the spatial attention residual module within the aforementioned CAM-U-Net segmentation model. The two are completely different in terms of processing dimension, input data format, operational location, and functional objectives. Specifically: the spatial attention residual module processes two-dimensional space (…). The input consists of a multi-channel 2D feature map and a single-channel CAM sub-heatmap, applied to each layer of the U-Net decoder. Its function is to perform soft gating on spatial location, enhancing the target region and suppressing the background, outputting an enhanced multi-channel 2D feature map. The gating attention fusion module in this stage processes a one-dimensional feature vector (channel dimension). Its input consists of two one-dimensional feature vectors (from the global stream and the local stream, respectively), applied after the final feature vectors from the dual-branch outputs are concatenated and before the classification layer. Its function is to dynamically assign weights to the global and local feature channels, determining the contribution ratio of the two types of information, outputting a fused one-dimensional feature vector.

[0071] As can be seen, the spatial attention residual module is used to implement pixel-level segmentation tasks by applying position-related attention in the two-dimensional image space; the gated attention fusion module is used to implement object-level classification tasks by applying global / local modality-related attention on the one-dimensional feature vector.

[0072] The specific workflow of the gated attention fusion module is as follows: (1) Transform the one-dimensional global feature vector With one-dimensional local eigenvectors By concatenating along the channel dimension, a fused concatenated feature vector is obtained. .

[0073] (2) The input consists of an attention subnet comprising two fully connected (FC) layers and a ReLU activation function. The first FC layer reduces the input dimension to half of its original dimension, and the second FC layer outputs a dimension of 1. The computation process of this attention subnet is represented as follows: ; in, For learnable weight matrix, For bias, Use the Sigmoid activation function; output It is a scalar with a value between [0,1], which is the gating coefficient.

[0074] (3) Using this gating coefficient As a soft switch, the two sets of features are dynamically recombined to obtain a fused feature vector: ; Among them, when When the value approaches 1, for example when cervical spine abnormalities or overall soft tissue morphology are the main risks, the model relies more on global features; when... When the value is close to 0, for example when the narrowest cross section of the airway or the volume of the tongue root are the main influencing factors, the model relies more on local features; thus, the mechanism can adaptively adjust the weights of global and local features.

[0075] (4) Fuse feature vectors The data is fed into the Softmax classification layer, which outputs the probability value of belonging to the "difficult airway" category, with a value range of [0,1], thereby achieving a fast and accurate judgment from the original image to the final result.

[0076] It is evident that, on the one hand, this gated attention fusion module differs from conventional multimodal fusion methods in its simple concatenation or fixed weighted averaging approach. By employing a learnable gating mechanism, the gating coefficients are automatically adjusted based on data. On the other hand, this mechanism differs from existing feature-level attention mechanisms, specifically targeting modality-level global / local fusion and acting on the top-level feature vector of the dual-branch output, complementing the spatial attention within the network. Furthermore, this module contains only two fully connected layers, has a very small number of parameters, and uses scalar gating coefficients, resulting in extremely low computational cost and facilitating end-to-end training. Thus, it achieves adaptive fusion of global anatomical and local morphological features, improving the accuracy of binary classification predictions for difficult airways.

[0077] In some implementations, for training the 3D SE-ResNet-18 prediction model in the difficult airway prediction and assessment module, the embodiments of the present invention designed the following three loss functions to address the problems of class imbalance and classification difficulties in clinical samples; including: (1) Global loss: ; In real-world clinical scenarios, the number of normal airway (negative) samples far exceeds the number of difficult airway (positive) samples. Without constraints, the model may tend to predict all samples as normal, thus achieving high overall accuracy but sacrificing its ability to identify difficult airways. Therefore, a class balance factor α is introduced. When there are fewer difficult airway samples, the value of α is increased to ensure the model gives greater attention to difficult airway features. In this loss function, y represents the true label (takes a value of 0 or 1), and p represents the probability that the model predicts a sample belongs to the positive category (i.e., difficult airway).

[0078] (2) Local loss: Since the soft probability ROI mask sequence is quite important for model prediction, in addition to maintaining the global loss... In addition to the loss function structure, this embodiment also introduces a focusing parameter. This method is used to handle samples with ambiguous classification boundaries and high recognition difficulty. The formula for local loss is as follows: ; in, This represents the focus adjustment factor (initially set to 2.0); for label 1, when When the value is close to 1 (easily classifiable samples). When the weight is close to 0, the weight is small; when When the value is close to 0 (difficult-to-classify samples). A value close to 1 indicates a relatively high weight.

[0079] (3) Total loss: ; in, The scaling factor, located between [0,1], is used to balance the weights of global and local losses, enabling the entire dual-branch prediction model to converge smoothly under a unified objective.

[0080] Furthermore, during the training phase, the sample size required for the 3D SE-ResNet-18 prediction model was estimated based on the area under the receiver operating characteristic curve (ROC-AUC). For example, a two-sided z-test was used with a significance level of 0.05 and a power of 80%. It was estimated that 795 samples needed to be included, including 159 cases in the difficult laryngoscopy group and 636 cases in the easy laryngoscopy group, to detect an AUC difference of 0.074 (previous result was 0.776, target value was 0.850). Considering a potential loss to follow-up rate of 20% and reserving 20% ​​of the sample for cross-validation, a final sample size of 1113 subjects was determined. If the actual sample size is insufficient during the modeling process, the sample size can be increased as needed.

[0081] Meanwhile, the training of the aforementioned prediction model was performed on a GPU using the PyTorch deep learning framework, employing the AdamW optimizer with a batch size of 8, a learning rate of 5e-5, a weight decay of 0.05, and a dropout rate of 0.3 before the fully connected classification layer to suppress overfitting. The training strategy consisted of 100 complete training cycles including 5 linear warm-ups, followed by a stepped decay strategy for the learning rate (multiplying the learning rate by a decay factor of 0.8 every 10 cycles).

[0082] Furthermore, to evaluate the performance of the predictive model, this embodiment employs a stratified five-fold cross-validation method. The dataset is first randomly shuffled to eliminate sequence bias, and then divided into five distinct subsets. During the partitioning process, the stratified sampling principle is strictly followed to ensure that the ratio of "difficult" to "non-difficult" airway samples in each subset remains consistent with that of the original dataset.

[0083] In each iteration of the cross-validation process, one subset is selected in turn as the test set specifically for model validation and performance evaluation, while the remaining four subsets are used as the training set. This process is repeated five times, with each subset serving as the test set exactly once. The average performance metric from the five experiments is then reported, enabling robust performance estimation across the entire dataset. After model training and validation are complete, the model is deployed in a real-world application environment.

[0084] To further illustrate the difficult airway assessment system of the present invention that utilizes MRI images combined with deep learning algorithms, a specific embodiment is provided below to explain the process of difficult airway prediction and assessment.

[0085] First, head and neck MRI scan sequences were acquired from the subject to be evaluated, resulting in 12-13 consecutive raw two-dimensional MRI image slice sequences (size: (pixels, grayscale image).

[0086] Second, each original slice is sequentially input into the pre-trained CAM-ResNet-18 generative model to generate four CAM saliency heatmaps (single-channel, size) corresponding to each original slice. The pixels are continuously probabilityd in the range [0,1], and the four heatmaps are weighted and fused into a final saliency heatmap based on the optimized hyperparameters.

[0087] Third, the final saliency heatmap is input into the trained CAM-U-Net segmentation model, which outputs a three-channel soft probability ROI mask (size) for the original slice. The three channels correspond to the upper airway, the base of the tongue, and the epiglottis, respectively. The above steps are repeated for all the original slices of the object to be evaluated to obtain the soft probability ROI mask sequence of the three channels.

[0088] Fourth, the original two-dimensional MRI image slice sequences are stacked along the Z-axis into a three-dimensional voxel matrix (size...). , (Number of slices), and simultaneously, the corresponding soft probability ROI mask sequence is compressed into a single channel by taking the maximum value of three channels on each slice, and then stacked along the Z-axis to form a three-dimensional mask volume (size). ).

[0089] Fifth, the aforementioned 3D voxel matrix is ​​input into the global flow of the dual-branch 3D SE-ResNet-18 prediction model, and the 3D mask volume is input into the local flow. After feature extraction, the gating coefficients of the two branches are dynamically calculated through a gating attention fusion module. The system outputs a fused feature vector, which is then processed by a Softmax classification layer to output the probability value that the object being evaluated belongs to the "difficult airway" category. .

[0090] Sixth, based on the preset decision threshold (set to 0.5), if If the airway is not difficult, it is classified as a difficult airway; otherwise, it is classified as a non-difficult airway. Simultaneously, the system presents the prediction results, along with a visual heatmap of the intermediate process and a segmentation mask, to clinicians as supplementary diagnostic evidence.

[0091] Therefore, the entire processing flow, from inputting the original two-dimensional MRI image slice sequence to outputting the predicted assessment results, can be completed within 30 seconds for a single subject, meeting the needs of real-time clinical auxiliary diagnosis. It should be noted that this system is a computer-aided analysis tool; the predicted assessment results for difficult airways are only for reference by clinicians and cannot replace the professional judgment of physicians.

[0092] It is understood that the above embodiments are only for ease of understanding and simplification of description, and should not be construed as limitations on the present invention. The present invention does not specifically limit the types of key soft tissue regions, the CAM-U-Net segmentation model, the construction and execution methods of the three-dimensional SE-ResNet-18 prediction model, etc.

[0093] Therefore, the embodiments of the present invention, on the one hand, based on the head and neck MRI images of the subject to be evaluated, can automatically identify areas with blurred soft tissue edges and large morphological variations using saliency heatmaps, and output soft probability masks corresponding to key soft tissue areas such as the upper airway, tongue base, and epiglottis, effectively overcoming problems such as difficulties in soft tissue quantification, insufficient segmentation accuracy, and manual measurement bias. On the other hand, through a weight-sharing dual-branch 3D network, global anatomical features (such as the overall topology of the cervical spine and mandible) and local geometric features (such as the narrowest cross section of the airway and the volume of the tongue base) can be extracted separately. The contribution ratio of the two types of features is adaptively adjusted through a gated attention fusion module, achieving deep fusion of global structure and local details, avoiding the limitations of a single feature source, and significantly improving the effect and accuracy of difficult airway prediction.

[0094] In another embodiment of the present invention, a computer device is provided, including at least one processor and at least one memory communicatively connected to said processor; The memory stores instructions executable by the processor to implement a difficult airway assessment system using MRI images combined with deep learning algorithms as described in any of the foregoing embodiments.

[0095] It should be noted that all processing procedures described in this system are executed by the aforementioned computer equipment. The core of this invention lies in using a difficult airway assessment system that combines MRI images with deep learning algorithms to analyze, perform mathematical operations, and map features onto acquired MRI image slices. The output predictive assessment results are auxiliary information. This system and equipment do not directly perform any diagnostic or treatment operations on living human bodies, nor do they output final clinical diagnostic conclusions. Instead, they provide an information analysis method based on data processing. This system does not replace the clinical judgment of physicians. Therefore, the technical solution of this invention belongs to computer-implemented information processing methods.

[0096] The above system and device embodiments are based on the same principles, and their related aspects can be referenced from each other to achieve the same technical effects. For specific implementation processes, please refer to the foregoing embodiments, which will not be repeated here.

[0097] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0098] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A difficult airway assessment system utilizing MRI images combined with deep learning algorithms, characterized in that, include: The heatmap generation module is used to obtain the original two-dimensional MRI image slice sequence of the object to be evaluated and generate a saliency heatmap corresponding to each original slice. The salient thermograms are used to indicate at least the key soft tissue areas of the upper airway, the base of the tongue, and the epiglottis; The mask image segmentation module is used to generate a soft probability ROI mask sequence based on the saliency heatmaps and the original two-dimensional MRI image slice sequence. Each mask in the soft probability ROI mask sequence is a multi-channel probability map, and different channel probability maps correspond to the confidence probability of the upper airway, the root of the tongue and the epiglottis region, respectively. The difficult airway prediction and assessment module is used to output the difficult airway prediction and assessment results of the subject under assessment based on the original two-dimensional MRI image slice sequence and the corresponding soft probability ROI mask sequence.

2. The difficult airway assessment system utilizing MRI images combined with deep learning algorithms according to claim 1, characterized in that, The heatmap generation module is implemented based on the CAM-ResNet-18 generation model, which integrates the CAM class activation mapping algorithm and the ResNet-18 residual network. The CAM-ResNet-18 generation model includes at least: shallow residual blocks for outputting feature maps with high spatial resolution and low channel number, and deep residual blocks for outputting feature maps with low spatial resolution and high channel number.

3. The difficult airway assessment system utilizing MRI images combined with deep learning algorithms according to claim 2, characterized in that, The heatmap generation module includes: a saliency heatmap generation unit; the saliency heatmap generation unit is used for: Based on the shallow residual blocks and deep residual blocks, feature extraction is performed on the original two-dimensional MRI image slice sequence, and multi-dimensional feature maps with different spatial resolutions and channel numbers corresponding to a single original slice are obtained. Based on the CAM-type activation mapping algorithm, multiple initial heatmaps of different resolutions are generated for each original slice's corresponding multidimensional feature map. Each initial heatmap corresponding to each original slice is upsampled and unified to the original resolution corresponding to each original slice to obtain multiple upsampled heatmaps. The multiple upsampled heatmaps are weighted and summed according to preset fusion weights to obtain the saliency heatmap corresponding to each original slice.

4. The difficult airway assessment system utilizing MRI images combined with deep learning algorithms according to claim 3, characterized in that, The heatmap generation module further includes: an evaluation and verification unit for determining whether the saliency heatmap satisfies the dual verification path; The dual verification path includes: Qualitative verification is used to overlay each of the aforementioned saliency heatmaps with their corresponding original slices in a semi-transparent pseudo-color manner, and based on visual examination information input by clinical experts, to determine whether the key soft tissue areas in the overlaid images match the actual anatomical locations of the tongue base, epiglottis, and upper airway. Quantitative evaluation is used to compare the significance heatmap corresponding to each original slice with the discrete binary mask manually annotated by clinical experts to determine whether the comparison result meets the preset threshold condition. The evaluation and verification unit is further configured to transmit the saliency heatmap obtained through dual evaluation to the mask image segmentation module after both the qualitative verification and the quantitative evaluation have passed the evaluation.

5. The difficult airway assessment system utilizing MRI images combined with deep learning algorithms according to claim 3, characterized in that, The mask image segmentation module is implemented based on an enhanced CAM-U-Net segmentation model that integrates the CAM class activation mapping algorithm and the U-Net network; wherein, the CAM-U-Net segmentation model includes at least: a U-Net encoder for extracting multi-scale feature maps, a U-Net decoder with different spatial resolution levels, and a spatial attention residual module for performing spatial attention enhancement processing.

6. The difficult airway assessment system utilizing MRI images combined with deep learning algorithms according to claim 5, characterized in that, The mask image segmentation module includes at least: a heatmap prior pyramid unit and a feature stitching unit; The heatmap prior pyramid unit is used to: use the saliency heatmap corresponding to each original slice in the original two-dimensional MRI image slice sequence as the bottom layer of the pyramid; use the bilinear interpolation algorithm to dynamically downsample each saliency heatmap to generate sub-heatmaps that match the feature map size of each level of the U-Net decoder, so as to form the heatmap prior pyramid; The feature stitching unit is used to: stitch the feature map of the U-Net decoder after upsampling at the previous level with the skip connection feature map of the corresponding level of the U-Net encoder along the channel dimension to obtain the original decoding features of each layer; introduce a matching sub-heatmap at each level of the U-Net decoder, and input each sub-heatmap into the spatial attention residual module to fuse it with the original decoding features of the corresponding level to obtain a multi-channel stitched feature map.

7. The difficult airway assessment system utilizing MRI images combined with deep learning algorithms according to claim 6, characterized in that, The spatial attention residual module is also used to perform the following operations on each level of the U-Net decoder: Receive the multi-channel stitched feature map corresponding to the current level and the sub-heatmap that matches the spatial resolution of the level; The sub-heatmap is input into a convolutional layer and processed by a Sigmoid activation function to generate a spatial attention weight map. The spatial attention weight map is broadcast by channel and then multiplied element-wise with the multi-channel stitched feature map to obtain the enhanced feature map; The enhanced feature map is added element-wise to the multi-channel stitched feature map to output the final feature map of this level. The final feature map is then convolved to output a soft probability ROI mask. The soft probability ROI mask sequence is obtained based on the soft probability ROI mask corresponding to each original slice.

8. The difficult airway assessment system utilizing MRI images combined with deep learning algorithms according to claim 7, characterized in that, The difficult airway prediction and evaluation module is implemented based on a three-dimensional SE-ResNet-18 prediction model that integrates the SE compression-excitation channel attention mechanism with the ResNet-18 residual network. The three-dimensional SE-ResNet-18 prediction model includes at least: a first branch for extracting global anatomical features, a second branch for extracting local anatomical features, a gated attention fusion module for adaptively fusing the output features of the two branches, and a Softmax classification layer.

9. The difficult airway assessment system utilizing MRI images combined with deep learning algorithms according to claim 8, characterized in that, The difficult airway prediction and assessment module is used for: Based on the first branch, the three-dimensional voxel matrix obtained by re-stacking the original two-dimensional MRI image slice sequence of the object to be evaluated along the physical space order is used as input to extract global anatomical features and output a one-dimensional global feature vector. Based on the second branch, the three-dimensional soft probability mask volume obtained by re-stacking the soft probability ROI mask sequence of the object to be evaluated along the physical space order is used as input to extract local morphological features and output a one-dimensional local feature vector. Based on the gated attention fusion module, the one-dimensional global feature vector and the one-dimensional local feature vector are concatenated along the channel dimension to obtain a concatenated feature vector; a gating coefficient is generated based on the concatenated feature vector, and the one-dimensional global feature vector and the one-dimensional local feature vector are weighted and summed using the gating coefficient to obtain a fused feature vector; the fused feature vector is input to the Softmax classification layer to output the classification prediction probability of the difficult airway, so as to obtain the prediction evaluation result of the difficult airway.

10. A computer device, characterized in that, It includes at least one processor and at least one memory communicatively connected to the processor; The memory stores instructions executable by the processor to implement the difficult airway assessment system using MRI images combined with deep learning algorithms as described in any one of claims 1 to 9.

Images