A network compression system for medical image features

Abstract

The application belongs to the technical field of medical equipment, and particularly relates to a network compression system for medical image features, which comprises a central processing unit module, a power supply module, a shooting module, an operation module and a comparison module. The central processing unit module and the power supply module are connected through wires. The central processing unit module and the shooting module are connected through signals. The shooting module and the operation module are connected through signals. The operation module and the comparison module are connected through signals. The comparison module and the central processing unit module are connected through signals. The installation of a pair of grips can facilitate the patient to hold the grips. The grips can block the patient on both sides, reduce the possibility of falling off the support platform, and the patient can stabilize himself through the grips, thereby increasing the definition of the subsequent scanning image.

Description

Technical Field

[0001] This invention belongs to the field of medical equipment technology, specifically a network compression system for medical image features. Background Technology

[0002] With the development of medical imaging equipment, more and more medical images are becoming an important basis for doctors to diagnose diseases. By comparing multiple sets of image information, the patient's condition can be observed and corresponding treatment plans can be proposed.

[0003] Currently, medical images are often produced by medical scanning equipment, such as CT scanners, which use precisely collimated X-ray beams, gamma rays, and ultrasound to perform cross-sectional scans of a part of the human body.

[0004] In use, CT scanners vary in size due to the limitations of different scenarios. The support platform for patients to lie flat is relatively narrow, which may cause the patient to roll over and fall, resulting in injury and secondary harm. Therefore, a network compression system for medical image features is proposed to address the above problems. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies and address the challenges of using CT scanners, which vary in size due to limitations imposed by different environments, and where the support platform for lying down is often narrow, there is a risk of the patient flipping over and falling, potentially causing injury or secondary harm. Therefore, this invention proposes a network compression system for medical image features.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: A network compression system for medical image features, comprising a central processing unit (CPU) module, a power supply module, an imaging module, a computing module, and a comparison module; the CPU module and the power supply module are connected by wires; the CPU module and the imaging module are connected by signals; the imaging module and the computing module are connected by signals; the computing module and the comparison module are connected by signals; and the comparison module and the CPU module are connected by signals. The computation module is used to: measure the similarity between the preprocessed image and the feature images of the pooling layer of each node of the network using the peak signal-to-noise ratio method to determine the depth of the transfer network, and compress the network based on the similarity measurement results of the features between different channels within the same layer and the features between different layers using the peak signal-to-noise ratio method.

[0007] Preferably, the power supply module can supply power to the central processing unit module. When scanning the patient, the central processing unit module can issue commands to enable the imaging module to take multiple images of the patient. After the images are taken, the imaging module uploads the obtained information to the computing module. The computing module processes the images using a medical image classification network compression method based on feature similarity measurement. The processed images are then uploaded to the comparison module. The patient's condition is determined based on multiple image comparisons, and the information is then uploaded to the central processing unit module for medical staff to view.

[0008] Preferably, the medical image classification network compression method first performs adaptive image denoising and contrast adjustment on the medical images, then measures the similarity between the preprocessed images and the feature images of the pooling layers of each node in the network to determine the depth of the transfer network. Next, the peak signal-to-noise ratio method is used to measure the feature similarity between different channels within the same layer and between different layers, and the features are sorted according to the measurement results. Then, the network is compressed using the feature measurement results and the expected compression coefficient. Finally, the classification performance is verified using a medical image dataset. This method can effectively compress the network with a slight reduction in model performance.

[0009] Preferably, the imaging module includes a base; the CT scanner body is provided on the side wall of the base; a support platform is slidably connected to the top surface of the base; a first groove is opened on the top surface of the support platform; a handle is provided inside the first groove; during operation, this step utilizes the installation of a pair of handles to facilitate the patient's hand gripping the handles, thereby reducing the possibility of the patient falling off the support platform by blocking the patient's sides with the handles, and at the same time, the patient can stabilize themselves by holding the handles, increasing the clarity of subsequent scan images.

[0010] Preferably, the handle is provided with a rubber sleeve; the handle and the rubber sleeve are connected by multiple sets of support rods; during operation, this step utilizes the pressure applied by the patient's hand to the rubber sleeve, which causes the rubber sleeve to contract at the point where it contacts the hand, making it fit snugly against the patient's hand, increasing the friction between the patient and the handle, while the support rods can support the rubber sleeve to accommodate patients of different body types, increasing the adaptability of the equipment during operation.

[0011] Preferably, a water bladder is fixedly connected between the rubber sleeve and the grip sidewall; multiple sets of rubber membranes are fixedly connected to the rubber sleeve sidewall; the water bladder and the rubber membranes are connected by a connecting tube; during operation, this step utilizes the pressure applied by the patient's hand to the rubber sleeve, which causes the rubber membrane near the hand sidewall to expand when the rubber sleeve contracts, increasing the wrapping effect of the rubber membrane on the hand, and causing the hand to adhere to the sidewall of the rubber sleeve, reducing the possibility of slippage.

[0012] Preferably, a slide rail is fixedly connected inside the first slide groove; the handle slides above the slide rail; a fixed rotating shaft is fixedly connected to the bottom end of the handle; the fixed rotating shaft rotates on the side wall of the first slide groove; during operation, this step utilizes the rotatability of the fixed rotating shaft to retract and extend the handle inside the first slide groove according to the different body shapes of the patients, reducing obstruction to the patients and improving the adaptability of the equipment during operation.

[0013] Preferably, a second sliding groove is formed through the top surface of the grip; a limiting rod is slidably connected inside the second sliding groove; the second sliding groove and the limiting rod are connected by a spring; multiple sets of third sliding grooves are formed on the top surface of the slide rail; the bottom end of the limiting rod slides inside the third sliding groove; a roller is rotatably connected to the bottom end of the limiting rod; during operation, this step utilizes the fixing effect of the limiting rod on the grip to reduce the possibility of the grip slipping during the scanning process, which could cause the patient's hand to shake and affect the scanning imaging effect. At the same time, the rotatability of the roller can help the limiting rod smoothly enter the slide rail, increasing the stability of the equipment during operation.

[0014] The beneficial effects of this invention are: 1. This invention provides a network compression system for medical image features. By using a pair of grips, the system facilitates the patient's hand gripping the grips. The grips provide resistance to the patient's sides, reducing the possibility of the patient falling from the support platform. At the same time, the patient can stabilize themselves by holding the grips, increasing the clarity of subsequent scanned images.

[0015] 2. This invention provides a network compression system for medical image features. By utilizing the pressure applied by the patient's hand to the rubber sleeve, the rubber sleeve and the area where they fit together can be contracted, making it fit snugly against the patient's hand and increasing the friction between the patient and the grip. At the same time, the support rod can support the rubber sleeve to accommodate patients of different body types, increasing the adaptability of the equipment during operation. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a system flowchart of Embodiment 1; Figure 2 This is a perspective view of Embodiment 1; Figure 3 This is a perspective view of the grip in Embodiment 1; Figure 4 This is a cross-sectional view of the grip in this embodiment one; Figure 5 Here is a flowchart of a network-based compressed classification system based on feature similarity metrics; Figure 6 The original image and the preprocessed image; Figure 7 This is the feature activation map of layer 11 and layer 18 of the network; Figure 8 This is the feature activation map of layer 26 and layer 33 of the network; Figure 9 This is the feature activation map of layer 41 and layer 48 of the network; Figure 10 This is the feature activation map of layer 55 and layer 62 of the network; Figure 11 The PSNR results are for layers 5 and 12 of the network. Figure 12 The PSNR results are for layers 20 and 27 of the network. Figure 13 The PSNR results are between layers 5 and 12, and layers 20 of the network. Figure 14 The PSNR results are for the inter-layer connections between layers 5 and 27, and between layers 12 and 20 of the network. Figure 15 The PSNR results are for the inter-layer connections between layers 12 and 27, and between layers 20 and 27 of the network. Figure 16 This is a perspective view of Example 2; Legend: 1. Base; 11. CT scanner body; 12. Support platform; 13. No. 1 slide rail; 14. Handle; 2. Rubber sleeve; 21. Support rod; 3. Water bladder; 31. Rubber membrane; 3. Connecting tube; 4. Fixed rotating shaft; 41. Slide rail; 5. No. 2 slide rail; 51. Limiting rod; 52. No. 3 slide rail; 6. Roller; 7. Anti-slip strip. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 Please see Figure 1-15As shown, a network compression system for medical image features includes a central processing unit (CPU) module, a power supply module, an imaging module, a computing module, and a comparison module. The CPU module and the power supply module are connected by wires; the CPU module and the imaging module are connected by signals; the imaging module and the computing module are connected by signals; the computing module and the comparison module are connected by signals; and the comparison module and the CPU module are connected by signals. The computation module is used to: measure the similarity between the preprocessed image and the feature images of the pooling layer of each node of the network using the peak signal-to-noise ratio method to determine the depth of the transfer network, and compress the network based on the similarity measurement results of the features between different channels within the same layer and the features between different layers using the peak signal-to-noise ratio method.

[0019] The power supply module provides power to the central processing unit (CPU) module. During patient scanning, the CPU module issues commands to enable the imaging module to take multiple images of the patient. After each image is taken, the imaging module uploads the information to the computing module. The computing module processes the images using a medical image classification network compression method based on feature similarity measurement. The processed images are then uploaded to the comparison module, which determines the patient's condition based on multiple image comparisons. Finally, the information is uploaded to the CPU module for medical staff to view.

[0020] The proposed medical image classification network compression method first performs adaptive image denoising and contrast adjustment on the medical images. Then, it measures the similarity between the preprocessed images and the feature images of the pooling layers of each node in the network to determine the depth of the transfer network. Next, it uses the peak signal-to-noise ratio method to measure the feature similarity between different channels within the same layer and between different layers, and sorts the features according to the measurement results. Then, it compresses the network using the feature measurement results and the expected compression coefficient. Finally, it uses a medical image dataset to verify the classification performance. This method can effectively compress the network with a slight reduction in model performance. Noise filtering employs adaptive noise reduction processing, as shown in Equations (I) and (II); ; ; local neighborhood pixels Indicates; pixel-level filter As shown in equation (Ⅲ); ; In formula (Ⅲ) The noise variance is given. Adaptive noise filtering can effectively remove image noise and improve network classification performance, while contrast enhancement can enrich the grayscale information of the image and extract its intrinsic features. The stretching process is shown in equation (Ⅳ). ; The feature similarity measurement method adopts the peak signal-to-noise ratio method; PSNR is a pixel-level evaluation method that measures the similarity of feature images by calculating the difference between pixels of different feature images; the calculation process is shown in equations (V) and (VI); ; ; Where MSE represents the mean square error of the two contrasting feature images. and These are the height and width of the feature image, respectively. The number of bits per pixel. Indicates peak signal-to-noise ratio. The higher the value, the higher the similarity between the two images.

[0021] The transfer network approach effectively leverages the advantages of the existing network while enabling rapid and sufficient training with limited image samples. Addressing the challenges of network depth determination and internal redundant information compression when transferring a network to a new medical image domain, particularly the significant impact of multiple channel aggregation on network performance, a method based on medical image feature similarity measurement is proposed. First, image noise is filtered and contrast is adjusted in the medical images. The peak signal-to-noise ratio (PSNR) is used to measure the similarity between the preprocessed image and the pooling layers of the network, determining the transfer network depth. Second, considering intra-layer and inter-layer redundant features, feature similarity is measured for features within the same layer and between different layers, and the results are sorted. Third, channel pruning is performed on channels with high feature similarity to compress the network using the feature measurement results and the expected compression coefficient. Finally, classification is performed using a medical image dataset to verify the network compression effect. A network compression classification system based on feature similarity measurement is shown below. Figure 5 As shown.

[0022] The experiment used a dataset containing 5856 pediatric chest cavity images, with diagnostic results categorized as normal or pneumonia. Of these, 3883 were pneumonia images and 1349 were normal images used for training. 390 were pneumonia images and 234 were normal images used for testing. The chest cavity images were selected from retrospective studies of pediatric patients aged 1 to 5 years. To improve the dataset's authority, all chest cavity images were first screened to remove low-quality or unreadable images, ensuring image quality. Two specialist physicians then graded the diagnoses of the images. Finally, to reduce diagnostic errors, a third expert reviewed the test set.

[0023] To verify the effectiveness of the algorithm, simulation experiments were conducted on the dataset. The experimental environment consisted of Matlab R2021b, an i9-10900K CPU, an NVIDIA GeForce RTX 3080Ti GPU, 64GB of memory, and a 64-bit Windows 10 operating system.

[0024] The model evaluation metrics include five indicators: accuracy, precision, sensitivity, specificity, and F1 score. TP, TN, FP, and FN are defined as true positives, true negatives, false positives, and false negatives, respectively. The classification accuracy, precision, sensitivity, specificity, and F1 score are defined as shown in equations (VII) to (XI).

[0025] ; ; ; ; ; Original image and preprocessed image as follows Figure 5 .like Figure 5 As shown, compared with the original image, the preprocessed image has reduced noise and stretched contrast. The preprocessed image also has richer texture, edge, and contour features. Figures 7 to 10 These refer to the Fire-concat layers in the SqueezeNet network, specifically layers 11, 18, 26, 33, 41, 48, 55, and 62. Figures 7 to 10As shown, the image information displayed by the feature activation maps of the Fire-concat node pooling layer gradually decreases with increasing network depth. Therefore, for medical image classification, a deeper network does not necessarily mean a stronger feature extraction capability; an appropriate network depth is beneficial for network compression and acceleration. As shown in Table 2, the PSNR metric of the 41st network node pooling layer is significantly lower than that of the previous 33rd layer. The network depths of the 41st and 33rd layers are 47 and 39 layers, respectively. Table 3 shows the classification results of pneumonia images at different depths. As shown in Table 3, when the network depth reaches 39 layers, four of the five evaluation metrics reach their highest values, indicating that increasing network depth does not necessarily improve the classification performance of medical images. Finally, the total depth of the original SqueezeNet network after migration is determined to be 39 layers, namely the first 33 layers and the last 6 layers. The last 6 layers are the dropout layer, conv layer, relu layer, pool layer, prob layer, and output layer in the original SqueezeNet network.

[0026] Table 1. Information on node merging layers and backbone network convolutional layers ; Table 2. PSNR similarity results between the preprocessed image and the feature maps of each channel of the Fire-concat network node pooling layer. ; Table 3. Classification results of pneumonia images at different depths. ; Table 4. Classification results of pneumonia images under different compression ratios. ; Intra-layer feature similarity was measured for the Fire-SqueezeNet convolutional layers (layers 5, 12, 20, and 27) of the transferred network. The results are as follows: Figure 11 , Figure 12 As shown. Furthermore, inter-layer feature similarity was measured for layers 5, 12, 20, and 27 of the network, and the results are as follows. Figure 13 , Figure 14 , Figure 15 As shown in Table 4, the feature measurement results within the same layer and between different layers are sorted. Channel pruning and compression are performed on channels with high feature similarity by setting a compression coefficient. The various classification evaluation indicators after compression are shown in Table 4. According to Table 4, as the network compression coefficient increases, the network's classification performance gradually decreases, but the decrease is not significant. Therefore, using the PSNR method to measure the feature similarity of the network feature maps can effectively serve as the basis for channel pruning, verifying the effectiveness of the method.

[0027] The imaging module includes a base 1; a CT scanner body 11 is mounted on the side wall of the base 1; a support platform 12 is slidably connected to the top surface of the base 1; a first groove 13 is opened on the top surface of the support platform 12; a handle 14 is provided inside the first groove 13; during operation, when a patient needs a CT scan in the hospital, medical staff can accompany the patient to the base 1, and the patient can lie flat on the support platform 12, holding the handle 14 to stabilize themselves. At this time, the support platform 12 moves, bringing the patient into the center of the CT scanner body 11 for scanning. This step utilizes the installation of a pair of handles 14, which makes it convenient for the patient to hold the handles 14. The handles 14 act as a barrier on both sides of the patient, reducing the possibility of the patient falling off the support platform 12. At the same time, the patient can stabilize themselves by holding the handles 14, increasing the clarity of the subsequent scan images.

[0028] The grip 14 is provided with a rubber sleeve 2 on the outside; the grip 14 and the rubber sleeve 2 are connected by multiple sets of support rods 21; during operation, when the patient grips the grip 14, it will first come into contact with the rubber sleeve 2. At this time, under the support of the support rods 21, the rubber sleeve 2 will cause part of the rubber sleeve 2 to contract at the contact point with the hand, fitting the hand. This step uses the pressure applied by the patient's hand to the rubber sleeve 2, which can cause the rubber sleeve 2 to contract at the contact point with the hand, making it fit the patient's hand, increasing the friction between the patient and the grip 14. At the same time, the support rods 21 can support the rubber sleeve 2 to adapt to patients of different body types, increasing the adaptability of the equipment during operation.

[0029] A water bladder 3 is fixedly connected between the rubber sleeve 2 and the side wall of the handle 14; multiple sets of rubber membranes 31 are fixedly connected to the side wall of the rubber sleeve 2; the water bladder 3 and the rubber membranes 31 are connected by a connecting pipe 32; during operation, when the patient's hand comes into contact with the rubber sleeve 2 and is squeezed, the liquid inside the water bladder 3 will enter the rubber membrane 31 through the connecting pipe 32 under pressure, causing the rubber membrane 31 to expand, thereby increasing the contact area between the rubber membrane 31 and the patient's hand. This step utilizes the pressure applied by the patient's hand to the rubber sleeve 2, which can cause the rubber membrane 31 near the side wall of the hand to expand when the rubber sleeve 2 contracts, increasing the wrapping effect of the rubber membrane 31 on the hand, and causing the hand to adhere to the side wall of the rubber sleeve 2, reducing the possibility of slippage.

[0030] A slide rail 41 is fixedly connected inside the first slide groove 13; the handle 14 slides above the slide rail 41; a fixed rotating shaft 4 is fixedly connected to the bottom end of the handle 14; the fixed rotating shaft 4 rotates on the side wall of the first slide groove 13; during operation, when facing patients of different body types, medical staff can use the slide rail 41 to drive the handle 14 to slide inside the first slide groove 13. At the same time, the rotatability of the fixed rotating shaft 4 can be adjusted according to the patient's body type. This step utilizes the rotatability of the fixed rotating shaft 4 to adjust the handle 14 inside the first slide groove 13 according to the different body types of patients, reducing obstruction to the patient and improving the adaptability of the equipment during operation.

[0031] The top surface of the grip 14 has a second sliding groove 5; a limiting rod 51 is slidably connected inside the second sliding groove 5; the second sliding groove 5 and the limiting rod 51 are connected by a spring; the top surface of the slide rail 41 has multiple sets of third sliding grooves 52; the bottom end of the limiting rod 51 slides inside the third sliding groove 52; a roller 6 is rotatably connected to the bottom end of the limiting rod 51; during operation, after the medical staff determines the position of the grip 14, the limiting rod 51 will insert into the third sliding groove 52 under the action of the spring force, fixing the position of the grip 14. After the patient's scan is completed, the staff can pull the limiting rod 51 upwards, causing it to disengage from the third slide groove 52. This allows the fixed rotating shaft 4 to rotate, enabling the handle 14 to move. This step utilizes the fixing effect of the limiting rod 51 on the handle 14 to reduce the possibility of the handle 14 slipping during the scan, which could cause the patient's hand to shake and affect the scanning imaging effect. At the same time, the rotatability of the roller 6 can help the limiting rod 51 smoothly enter the slide rail 41, increasing the stability of the equipment during operation.

[0032] Example 2 Please see Figure 5 As shown in the first embodiment, as another implementation of the present invention, the rubber sleeve 2 is fitted with an anti-slip strip 7; the anti-slip strip 7 is arranged in a spiral downwards; during operation, when the patient holds the rubber sleeve 2, the presence of the rubber sleeve 2 can provide the patient with an additional contact area. This step, by using the installation of the rubber sleeve 2, can further improve the contact between the patient's hand and the rubber sleeve 2, increase friction, reduce hand slippage, and improve stability.

[0033] Working principle: During operation, when a patient needs a CT scan in the hospital, medical staff can accompany them to the base 1. The patient can lie flat on the support platform 12, holding the handle 14 to secure themselves. At this time, the support platform 12 moves, bringing the patient into the center of the CT machine body 11 for scanning. When the patient holds the handle 14, it will first contact the rubber sleeve 2. At this time, under the support of the support rod 21, the rubber sleeve 2 will cause part of the rubber sleeve 2 to contract at the contact point with the hand, fitting the hand. When the patient's hand contacts the rubber sleeve 2 and squeezes, the liquid inside the water bladder 3 will enter the rubber membrane 31 through the connecting tube 32 under pressure. The expansion of the rubber membrane 31 increases the contact area between the rubber membrane 31 and the patient's hand. When dealing with patients of different body shapes, medical staff can slide the handle 14 into the first groove 13 via the slide rail 41. At the same time, the rotatability of the fixed rotating shaft 4 can be adjusted according to the patient's body shape. After the medical staff determines the position of the handle 14, the limiting rod 51 will be inserted into the third groove 52 under the action of the spring force to fix the position of the handle 14. After the patient scan is completed, the staff can provide an upward pull-out effect for the limiting rod 51, causing the limiting rod 51 to disengage from the third groove 52, so that the fixed rotating shaft 4 can rotate and the handle 14 can move.

[0034] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0035] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A network compression system for medical image features, characterized in that: It includes a central processing unit (CPU) module, a power supply module, an imaging module, a computing module, and a comparison module; the CPU module and the power supply module are connected by wires; the CPU module and the imaging module are connected by signals; the imaging module and the computing module are connected by signals; the computing module and the comparison module are connected by signals; and the comparison module and the CPU module are connected by signals. The computation module is used to: measure the similarity between the preprocessed image and the feature images of the pooling layer of each node of the network using the peak signal-to-noise ratio method to determine the depth of the transfer network, and compress the network based on the similarity measurement results of the features between different channels within the same layer and the features between different layers using the peak signal-to-noise ratio method.

2. The network compression system for medical image features according to claim 1, characterized in that: The power supply module provides power to the central processing unit (CPU) module. During patient scanning, the CPU module issues commands to enable the imaging module to take multiple images of the patient. After each image is taken, the imaging module uploads the information to the computing module. The computing module processes the images using a medical image classification network compression method based on feature similarity measurement. The processed images are then uploaded to the comparison module, which determines the patient's condition based on multiple image comparisons. Finally, the information is uploaded to the CPU module for medical staff to view.

3. The network compression system for medical image features according to claim 2, characterized in that: The proposed medical image classification network compression method first performs adaptive image denoising and contrast adjustment on the medical images. Then, it measures the similarity between the preprocessed image and the feature images of the pooling layers of the network to determine the transfer network depth. Next, it uses the peak signal-to-noise ratio (PSNR) method to measure feature similarity between different channels within the same layer and between different layers, and sorts the results according to the measurement. Third, it compresses the network using the feature measurement results and the expected compression coefficient. Finally, it uses a medical image dataset for classification to verify performance. This method can effectively compress the network with a slight reduction in model performance. Adaptive image denoising, as shown in Equations (I) and (II); ; local neighborhood pixels Indicates; pixel-level filter As shown in equation (Ⅲ); ； In formula (Ⅲ) For noise variance; Adaptive noise filtering can effectively remove noise from images and improve network classification performance. Contrast stretching can enrich the grayscale information of images and extract the intrinsic features of images. The stretching process is shown in Equation (Ⅳ). ; The peak signal-to-noise ratio method is used to measure feature similarity. PSNR is a pixel-level evaluation method that measures the similarity of feature images by calculating the difference between pixels of different feature images; the calculation process is shown in equations (V) and (VI).

4. The network compression system for medical image features according to claim 3, characterized in that: The model evaluation metrics for the central processing module are accuracy, precision, sensitivity, specificity, and F1 score; TP, TN, FP, and FN are defined as true positives, true negatives, false positives, and false negatives, respectively. ； The classification accuracy, precision, sensitivity, specificity, and F1 score are defined as shown in equations (VII) to (XI).

5. A network compression system for medical image features according to claim 4, characterized in that: The system's processing method for the original and preprocessed images involves comparing them with the original image. Preprocessing reduces noise and enhances contrast, resulting in richer texture, edge, and contour features. The pseudocode for the medical image classification network compression method based on feature similarity measurement is as follows: Algorithm 1 Input: Original image data matrix H, compression coefficient a; Output: Confusion matrix F; (1) Preprocess image H to obtain A; (2) Plot A onto the pre-training grid, calculate the similarity between input A and each layer, and determine the grid depth d; (3) Calculate the intra-layer similarity and inter-layer similarity between feature images in the network; (4) Prune and compress the network according to the compression factor a; (5) Retrain the network to obtain the weight matrix W and the bias b; (6) For epoch=1: k; (7) Calculate the sample classification loss; (8) Update the weight matrix W' and the bias b'; (9) End for; (10) Repeat steps 4 to 10 until the network training under different compression coefficients is completed, and then fine-tune the network; (11) Obtain the confusion matrix F and five classification indicators.

6. A network compression system for medical image features according to claim 5, characterized in that: The imaging module includes a base (1); the side wall of the base (1) is provided with a CT machine body (11); the top surface of the base (1) is slidably connected to a support platform (12); the top surface of the support platform (12) is provided with a first slide groove (13); the first slide groove (13) is provided with a handle (14); the handle (14) is provided with a rubber sleeve (2); the handle (14) and the rubber sleeve (2) are connected by multiple sets of support rods (21); a water bladder (3) is fixed between the side wall of the rubber sleeve (2) and the handle (14); multiple sets of rubber membranes (31) are fixed to the side wall of the rubber sleeve (2); the water bladder (3) and the rubber membranes (31) are connected by a connecting pipe (32).

7. A network compression system for medical image features according to claim 6, characterized in that: A slide rail (41) is fixedly connected inside the first slide groove (13); the handle (14) slides above the slide rail (41); a fixed rotating shaft (4) is fixedly connected to the bottom end of the handle (14); the fixed rotating shaft (4) rotates on the side wall of the first slide groove (13); a second slide groove (5) is opened through the top surface of the handle (14); a limit rod (51) is slidably connected inside the second slide groove (5); the second slide groove (5) and the limit rod (51) are connected by a spring; multiple sets of third slide grooves (52) are opened on the top surface of the slide rail (41); the bottom end of the limit rod (51) slides inside the third slide groove (52); a roller (6) is rotatably connected to the bottom end of the limit rod (51).

Images