Laparoscopic lesion target detection method based on VITDET algorithm

Abstract

A laparoscopic lesion target detection method based on a VITDET algorithm can calculate and match the extracted key surgical elements and the surgical safety model during actual surgical operation, and can calculate in real time during the operation and prompt in the form of text and pictures, effectively assisting surgeons to improve surgical safety.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of medical computer science and artificial intelligence computer vision algorithm design, specifically to a method for laparoscopic lesion target detection based on VITDET-type algorithms. Background Technology

[0002] The medical device market is one of the fastest-growing and most actively traded markets in the global economy, and the laparoscopic market is one of the fastest-growing segments within the global medical device market. Although the laparoscopic market was previously mainly in Europe and North America, in recent years, with the improvement of China's economic level and the perfection of its medical system, the domestic medical device market has flourished, with a significant increase in demand for laparoscopic equipment.

[0003] Similar to gastroscopy and colonoscopy, a laparoscopy device is a medical device equipped with a light source and a miniature camera, classified as a Class III medical device. The complete system includes the laparoscope, energy system, light source system, irrigation system, imaging system, and commonly used laparoscopic surgical instruments (such as puncture needles, cannulas, dissecting forceps, electrocoagulation hooks and rods, clamp applicators, titanium clips, scissors, needle holders, snares, etc.), primarily used in general surgery, gynecological surgery, and other surgical procedures. During laparoscopic surgery, the surgeon inserts the laparoscope into the abdominal cavity through a small incision. The camera on the laparoscope transmits images of the internal abdominal cavity to a monitor on a video surveillance system for the surgeon's examination and surgical procedures. Compared to traditional open surgery, laparoscopic surgery offers advantages such as less trauma, less postoperative pain, faster recovery, and better aesthetic results. Therefore, it is widely favored by both doctors and patients in clinical practice.

[0004] In recent years, the field of computer vision (CV) in AI has developed rapidly, with object detection being one of the main research areas within CV. With the advancement of computer technology, artificial intelligence (AI) has been widely studied in the field of endoscopy, such as for the identification of abnormal tumor sites based on laparoscopy. A major role of computer-aided diagnosis is to help laparoscopists differentiate between neoplastic and non-neoplastic lesions, and multiple studies have demonstrated the potential of AI systems in the early qualitative diagnosis of masses. Furthermore, the diagnostic accuracy of AI systems has been compared with that of experienced physicians, proving to achieve comparable performance. Object detection-based algorithms can locate and annotate eight target areas of laparoscopic data, automatically select important targets, and promptly inform doctors of the extent of lesions.

[0005] However, current target detection algorithms for laparoscopic data still face several challenges. First, from an algorithmic perspective, current processing methods for laparoscopic data are still in their early stages, resulting in mediocre performance. Second, medical graphics often contain significant noise, making it impossible to use the raw data directly, yet only a limited amount of work addresses this noise in laparoscopic data. Data and models are the two key pillars of computer-aided medical tasks, and these issues directly hinder the application of computer vision in laparoscopy, necessitating further technological improvements. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a laparoscopic lesion target detection method based on the VITDET-type algorithm.

[0007] The technical solution adopted in this invention is a method for laparoscopic lesion target detection based on the VITDET algorithm, comprising the following steps:

[0008] Step 1: Collect videos and images of laparoscopic surgeries, and annotate the videos and images to obtain a laparoscopic detection database. The database includes training set data for training and test set data for testing.

[0009] Step 2: Use the histogram equalization algorithm to preprocess the data in the database;

[0010] Step 3: Resize the preprocessed data without distortion;

[0011] Step 4: Construct a target detection network for laparoscopic images based on VITDET-type algorithms. The specific steps include:

[0012] 4.1 Extracting data features using the Swin-Transformer backbone network of VITDET-type algorithms;

[0013] 4.2 Optimize the features obtained from the backbone network VIT using the attention mechanism to obtain more effective low-level features;

[0014] 4.3 Obtain deep branch features for the final detection task using the feature pyramid PANnet;

[0015] 4.4. Use YOLO Head to process the obtained features to obtain the final prediction results;

[0016] 4.5 Decode the prediction results and calculate the loss function;

[0017] 4.6 Draw the border of the detection result on the original image;

[0018] Step 5: Input the training set data obtained in Steps 1-3 into the object detection network to train the object detection network, obtain the object detection network model, and save it;

[0019] Step 6: Input the test set data obtained in Steps 1-3 into the target detection network model, and use the target detection network model to detect whether there are lesion target areas, and test the target detection network model;

[0020] Step 7: Deploy the object detection network model onto the GPU module to detect images transmitted by the client and return the detection results to the interactive interface.

[0021] Step 4.1 specifically includes the following steps:

[0022] 4.1.1 First, normalize the feature map;

[0023] 4.1.2. The shift_size parameter determines whether a shift operation is needed on the feature map;

[0024] 4.1.3 Divide the feature map into windows;

[0025] 4.1.4 Calculate attention, and use attn_mask to distinguish between Window Attention and ShiftWindow Attention, which is used to limit the content that each position in the attention can see;

[0026] 4.1.5. Merge all windows;

[0027] 4.1.6 If a shift operation was performed previously, reverse shift will be performed to restore the previous shift operation;

[0028] 4.1.7 Droppath and residual links;

[0029] 4.1.8 Then, a normalized fully connected layer, as well as droppath and residual connections are added.

[0030] In step 4.4, the YOLO Head consists of convolutional layers, max-pooling layers, and downsampling layers. Its feature processing specifically includes the following steps:

[0031] 4.4.1 Extract feature map branches with feature map sizes of 13×13, 26×26, and 52×52 from the backbone network as input to the YOLO Head;

[0032] 4.4.2 Perform convolutional pooling downsampling on these three branches, in the order of multiple processing structures from a depth of 13×13 to 52×52 and then back to a depth of 13×13;

[0033] 4.4.3 YOLO Head obtains the target category prediction results and detection box positions for the three branches through fully connected layer operations.

[0034] In step 4.5, the specific calculation method for the loss function is as follows:

[0035] 4.5.1 First, calculate the non-overlapping area S by summing the squares of all Euclidean distances between the corresponding distance values. The specific formula is as follows:

[0036] S=(L * -L) 2 +(T * -T) 2 +(R * -R) 2 +(B * -B) 2 ,

[0037] Where {L, T, R, B} and {L * T * R * B *} represents the difference between the center coordinates of the predicted and true labels' bounding boxes, respectively. L and T refer to the distances from the bottom right corner to the left and top boundaries, respectively, while R and B refer to the distances from the top left corner to the right and bottom boundaries, respectively.

[0038] The {L} of the real bounding box * T * R * B * The calculation formula is as follows:

[0039]

[0040]

[0041]

[0042] w (i) =w / s i ,

[0043] h (i) =h / s i ,

[0044] Where i represents the three different scales in this paper, S i ={8,16,32}, where (x,y) are the coordinates of the center point of the truth box. and Let i represent the coordinates of the top-left and bottom-right corners at point i. and Indicates the coordinates of the top left and bottom right corners of the center position;

[0045] The formula for calculating the {L, T, R, B} of the prediction box is as follows:

[0046]

[0047] Where σ represents the sigmoid function, p0, p1, p2, p3 represent the network's predictions of the distance values, and the sigmoid function is used to ensure that they are between [0,1], i = {1,2,4};

[0048] 4.5.2. The intersection region I is obtained by calculating the square of the length of the diagonal of the intersection region. The specific formula is as follows:

[0049] I = (w I ) 2 +(h I ) 2 ,

[0050] Among them, w I and h I These represent the width and height of the intersection area, respectively.

[0051] 4.5.3 Calculate the minimum region C covering the predicted and ground truth bounding boxes using the length and width. The specific formula is: C = (w... C ) 2 +(h C ) 2 ,

[0052] Among them, w C and h C The width and height of the minimum regions that respectively cover the predicted and ground truth bounding boxes, w C =max(L * ,L)+max(R * ,R)-1,h C =max(T) * ,T)+max(B * ,B)-1;

[0053] 4.5.4 By minimizing c, the predicted bounding box can move in four directions toward the ground truth bounding box, and the final SDIoU is calculated as follows: Where ρ represents the positive trade-off that benefits the overlapping region, the SDIoU loss is finally defined as loss = 1 - SDIoU.

[0054] In step 2, the specific steps of the histogram equalization algorithm are as follows:

[0055] 2.1. Scan each pixel of the original grayscale image sequentially and calculate the grayscale histogram of the image;

[0056] 2.2 Calculate the cumulative distribution function of the gray-level histogram;

[0057] 2.3. Obtain the mapping relationship between input and output based on the cumulative distribution function and histogram equalization principle;

[0058] 2.4 Finally, perform image transformation based on the results obtained from the mapping relationship.

[0059] In step 3, the method for resizing without distortion is as follows: First, scale the original image proportionally, and then add grayscale bars to the parts that do not meet the size requirements.

[0060] The beneficial effects of this invention are:

[0061] 1. During actual surgical procedures, the extracted key surgical elements and surgical safety models are calculated and matched. The calculations are performed in real time during the operation and prompts are presented in the form of pictures and text, effectively assisting surgeons in improving surgical safety.

[0062] 2. Construct a target detection network for laparoscopic images based on the VITDET-type algorithm. The VITDET-type algorithm is a superior target detection model that can achieve better target detection results, ensure accurate indication of anatomical location, guide surgical operation, indicate potential risks during operation, and reduce the probability of surgical damage.

[0063] 3. The color processing method based on histogram equalization algorithm makes the model more robust to the influence of lighting, can overcome the interference of color difference artifacts caused by the use of multi-source devices, can alleviate the endoscope photometric distortion in actual laparoscopic scenarios, such as changes in light brightness and light intensity, as well as red light, effectively reduce data noise, and effectively improve the detection performance of the detection model. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating an embodiment of the present invention. Detailed Implementation

[0065] A method for laparoscopic lesion target detection based on VITDET-type algorithms includes the following steps:

[0066] Step 1: Collect videos and images of laparoscopic surgeries, and use CVAT to annotate the videos and images to obtain a laparoscopic detection database. The database includes training set data for training and test set data for testing.

[0067] Step 2: Preprocess the data in the database using the histogram equalization algorithm. The specific steps are as follows:

[0068] 2.1. Scan each pixel of the original grayscale image sequentially and calculate the grayscale histogram of the image;

[0069] 2.2 Calculate the cumulative distribution function of the gray-level histogram;

[0070] 2.3. Obtain the mapping relationship between input and output based on the cumulative distribution function and histogram equalization principle;

[0071] 2.4 Finally, perform image transformation based on the results obtained from the mapping relationship.

[0072] The aforementioned data color processing method makes the model more robust to the influence of lighting, overcomes the interference of color difference artifacts caused by the use of multi-source devices, and can effectively reduce data noise and improve the detection performance of the detection model by mitigating endoscope photometric distortion in actual laparoscopic scenarios, such as changes in light brightness and intensity, as well as red light.

[0073] Step 3: Modify the size of the preprocessed data without distortion. The specific method is as follows: First, scale the original image proportionally, and then add grayscale bars to the parts that do not meet the size requirements.

[0074] By uniformly resizing the input images, the images can be smoothly fed into the subsequent deep neural network to complete the model's feature extraction. The above-mentioned distortion-free resizing method will not destroy the original data features and ensure that the original data is not deformed.

[0075] Step 4: Construct a target detection network for laparoscopic images based on VITDET-type algorithms. The specific steps include:

[0076] 4.1 Extracting data features using the Swin-Transformer backbone network of VITDET-type algorithms;

[0077] 4.2 Optimize the features obtained from the backbone network VIT using the attention mechanism to obtain more effective low-level features;

[0078] 4.3 Obtain deep branch features for the final detection task using the feature pyramid PANnet;

[0079] 4.4. Use YOLO Head to process the obtained features to obtain the final prediction result:

[0080] 4.5 Decode the prediction results and calculate the loss function;

[0081] 4.6 Draw the border of the detection result on the original image.

[0082] Step 4.1 specifically includes the following steps:

[0083] 4.1.1 First, normalize the feature map;

[0084] 4.1.2. The shift_size parameter determines whether a shift operation is needed on the feature map;

[0085] 4.1.3 Divide the feature map into windows;

[0086] 4.1.4 Calculate attention, and use attn_mask to distinguish between Window Attention and ShiftWindow Attention, which is used to limit the content that each position in the attention can see;

[0087] 4.1.5. Merge all windows;

[0088] 4.1.6 If a shift operation was performed previously, reverse shift will be performed to restore the previous shift operation;

[0089] 4.1.7 Droppath and residual links;

[0090] 4.1.8 Then, a normalized fully connected layer, as well as droppath and residual connections are added.

[0091] The data is fed into a ViT-based Swin-Transformer to extract features. Four stages are used to construct feature maps of different sizes. One stage first passes through a Linear Embedding layer, then repeatedly stacks SwinTransformer Blocks. The remaining three stages first pass through a Patch Merging layer for downsampling, then repeatedly stack SwinTransformer Blocks. The SwinTransformer Block contains two structures: one using a W-MSA structure and the other using a SW-MSA structure; the two SwinTransformer Blocks are used in pairs.

[0092] Furthermore, the Windows Multi-head Self-Attention (W-MSA) module is introduced to reduce computational cost, enabling the construction of deeper networks. The paired nature of W-MSA is also crucial; by adjusting the window offset, it allows communication between neighboring windows, thus resolving the issue of information exchange between different windows. Additionally, to prevent information confusion caused by merging different regions, masked MSA is used in the actual computation. By setting a mask to isolate information from other regions, each pixel is calculated only with pixels in its own region, thus solving this problem.

[0093] A target detection network for laparoscopic images based on the VITDET-type algorithm was constructed. The VITDET-type algorithm is a superior target detection model that can achieve better target detection results, ensure accurate indication of anatomical location, guide surgical operation, indicate potential risks during operation, and reduce the probability of surgical damage.

[0094] Steps 4.2 and 4.3 process the features of specific branches in the backbone network, including SSP and PANet. SSP, as the first part of the enhanced feature extraction network in VITDET-like networks, not only fuses features of more different resolutions to obtain more relevant information, but also greatly increases the receptive field, thereby separating the most significant contextual features from the backbone features. PANet, the second part of the enhanced feature extraction network in VITDET-like networks, includes a convolutional upsampling and downsampling mechanism, which, through repeated sampling, can obtain more effective features for detection.

[0095] In step 4.4, the YOLO Head consists of convolutional layers, max-pooling layers, and downsampling layers. Its feature processing specifically includes the following steps:

[0096] 4.4.1 Extract feature map branches with feature map sizes of 13×13, 26×26, and 52×52 from the backbone network as input to the YOLO Head;

[0097] 4.4.2 Perform convolutional pooling downsampling on these three branches, in the order of multiple processing structures from a depth of 13×13 to 52×52 and then back to a depth of 13×13;

[0098] 4.4.3 YOLO Head obtains the target category prediction results and detection box positions for the three branches through fully connected layer operations.

[0099] The role of YOLO Head is to perform final processing on the results of the three effective feature layers obtained from PANet, and obtain the prediction results of the VITDET-like algorithm network. The prediction results represent whether the prior box on each grid contains an object, the type of object, and the adjustment parameters of this prior box.

[0100] In step 4.5, the decoding process of the VITDET-type algorithm refers to the adjustment of the prior box, specifically adjusting the center coordinates and width / height coordinates of the prior box. This involves translating and scaling the prior box to obtain the true bounding box. The center coordinate adjustment involves the offset distance between the centers of the prior box and the predicted box. The specific calculation method for the loss function is as follows:

[0101] 4.5.1 First, calculate the non-overlapping area S by summing the squares of all Euclidean distances between the corresponding distance values. The specific formula is as follows:

[0102] S=(L * -L) 2 +(T * -T) 2 +(R * -R) 2 +(B * -B) 2 ,

[0103] Where {L, T, R, B} and {L * T * R * B *} represents the difference between the center coordinates of the predicted and true labels' bounding boxes, respectively. L and T refer to the distances from the bottom right corner to the left and top boundaries, respectively, while R and B refer to the distances from the top left corner to the right and bottom boundaries, respectively.

[0104] The {L} of the real bounding box * T * R * B * The calculation formula is as follows:

[0105]

[0106]

[0107]

[0108] w (i) =w / s i ,

[0109] h (i) =h / s i ,

[0110] Where i represents the three different scales in this paper, S i ={8,16,32}, where (x,y) are the coordinates of the center point of the truth box. and Let i represent the coordinates of the top-left and bottom-right corners at point i. and Indicates the coordinates of the top left and bottom right corners of the center position;

[0111] The formula for calculating the {L, T, R, B} of the prediction box is as follows:

[0112]

[0113] Where σ represents the sigmoid function, p0, p1, p2, p3 represent the network's prediction of the distance value, and i = {1, 2, 4}. It should be noted that the sigmoid function is added to prevent the actual center point coordinates from being moved out during the prediction process. That is, using sigmoid ensures that they are between [0, 1]. Normalizing the coordinates of the predicted box center point ensures that the center point of the true box can only be translated within its own grid. The width and height parameters are obtained by scaling the width and height of the prior box to obtain the width and height of the predicted box.

[0114] 4.5.2. The intersection region I is obtained by calculating the square of the length of the diagonal of the intersection region. The specific formula is as follows:

[0115] I = (w I ) 2 +(h I ) 2 ,

[0116] Among them, w I and h I These represent the width and height of the intersection area, respectively.

[0117] 4.5.3 Calculate the minimum region C covering the predicted and ground truth bounding boxes using the length and width. The specific formula is: C = (w... C ) 2 +(h C ) 2 ,

[0118] Among them, w C and h C The width and height of the minimum regions that respectively cover the predicted and ground truth bounding boxes, w C =max(L * ,L)+max(R * ,R)-1,h C =max(T) * ,T)+max(B * ,B)-1;

[0119] 4.5.4 By minimizing c, the predicted bounding box can move in four directions toward the ground truth bounding box, and the final SDIoU is calculated as follows: Where ρ represents the positive trade-off that benefits the overlapping region, the SDIoU loss is finally defined as loss = 1 - SDIoU.

[0120] The loss function of the entire VITDET-like algorithm consists of three parts: First, the regression loss, which is calculated only for positive samples using SDIOU; this step is only performed for prior boxes with corresponding ground truth bounding boxes. Second, the cross-entropy loss for determining whether the prior box contains an object is calculated; this applies to both positive and negative samples, with the first part being the cross-entropy loss for positive samples and the second part for negative samples. Finally, the loss for determining whether the prior box contains an object category is calculated; here, only the category confidence loss for positive samples is calculated. Overall, the total loss function is the regression loss, plus the loss for determining whether the prior box contains an object and the category loss.

[0121] Step 5: Input the training set data obtained in Steps 1-3 into the object detection network to train the object detection network, obtain the object detection network model, and save it.

[0122] Step 6: Input the test set data obtained in Steps 1-3 into the target detection network model, and use the target detection network model to detect whether there are lesion target areas, and test the target detection network model.

[0123] Step 7: Deploy the object detection network model onto the GPU module to detect images transmitted by the client and return the detection results to the interactive interface.

[0124] Using the above scheme, during actual surgical procedures, the extracted key surgical elements and surgical safety models can be calculated and matched in real time, with prompts provided in graphic and textual form, effectively assisting surgeons in improving surgical safety. This falls within the scope of patent protection.

Claims

1. A laparoscopic lesion target detection method based on a VITDET algorithm, characterized in that, The method comprises the following steps: Step 1, collect videos and pictures of laparoscopic surgery, and label the videos and pictures to obtain a laparoscopic detection database, wherein the data in the database include training set data for training and test set data for testing Step 2, pre-process the data in the database using a histogram equalization algorithm; Step 3, modify the size of the pre-processed data without distortion; Step 4, construct a laparoscopic image target detection network based on a VITDET algorithm, and the specific steps include: 4.1, extracting data features by using a backbone network Swin-Transformer of the VITDET algorithm; 4.2, optimizing the features obtained by the backbone network VIT by using an attention mechanism to obtain more effective bottom features; 4.3, obtaining deep branch features used for the final detection task through a feature pyramid PANnet; 4.4, processing the obtained features by using a YOLO Head to obtain the final prediction result; 4.5, decoding the prediction result and calculating a loss function; 4.6, drawing a detection result frame on the original image; Step 5, inputting the training set data obtained in steps 1-3 into the target detection network for training the target detection network, obtaining a target detection network model and saving the target detection network model; Step 6, inputting the test set data obtained in steps 1-3 into the target detection network model to detect whether there is a lesion target area through the target detection network model, and testing the target detection network model; Step 7, deploying the target detection network model to a GPU module for detecting pictures transmitted by a client, and returning the detection result to an interactive interface.

2. The laparoscopic lesion target detection method based on the VITDET class algorithm according to claim 1, characterized in that, Step 4.1 specifically includes the following steps: 4.1.1, first, norm the feature map; 4.1.2, determine whether the feature map needs to be shifted by using a shift_size parameter; 4.1.3, divide the feature map into windows; 4.1.4, calculate attention, and distinguish whether it is Window Attention or Shift Window Attention by using an attn_mask to limit the content that each position can see in the attention; 4.1.5, merge the windows; 4.1.6, if the shift operation is performed, perform reverse shift to restore the previous shift operation; 4.1.7, droppath and residual connection; 4.1.8, then, pass the feature map through a norm+full connection layer, and perform droppath and residual connection. 3.The laparoscopic lesion target detection method based on the VITDET class algorithm of claim 1, wherein, In step 4.4, the YOLO Head is composed of a convolution layer, a maximum pooling layer and a down-sampling layer, and the feature processing specifically includes the following steps: 4.4.1, extracting feature map branches with sizes of 13x13, 26x26 and 52x52 in the backbone network as inputs of the YOLO Head; 4.4.2, performing convolution, pooling and down-sampling operations on the three branches, and the processing structure is a multiple processing structure from 13x13 deep layer to 52x52 and then to 13x13 deep layer. 4.4.3, YOLO Head gets the target class prediction results and the detection box position of three branches through the full connection layer operation. 4.The laparoscopic lesion target detection method based on the VITDET class algorithm of claim 1, wherein, In step 4.5, the specific calculation method of the loss function is as follows: 4.5.1, first calculate the non-overlapping area S by adding the square of all Euclidean distances between the corresponding distance values, and the specific formula is: S = (L * - L) 2 + (T * - T) 2 + (R * - R) 2 + (B * - B) 2 , where {L, T, R, B} and {L * , T * , R * , B *} are the true difference of bounding box center point coordinates of prediction and ground truth, respectively, L and T refer to the distance from the bottom right corner to the left and top boundaries, and R and B refer to the distance from the top left corner to the right and bottom boundaries; The {L * , T * , R * , B *} calculation formula is as follows: w (i) = w / s i , h (i) = h / s i , where i denotes the three different scales in this paper, S i = {8, 16, 32}, (x, y) is the center point coordinate of the ground truth box, and denote the respective coordinates of the top-left and bottom-right corners at scale i, and denote the respective coordinates of the top-left and bottom-right corners of the center position; The calculation formula of the prediction box {L, T, R, B} is as follows: Where σ represents the sigmoid function, p0, p1, p2, p3 represent the network prediction of distance value, and sigmoid is used to ensure that they are between [0, 1], i={1, 2, 4}; 4.5.2, obtain the intersection area I by calculating the square of the length of the diagonal of the intersection area, and the specific formula is: I = (w I ) 2 + (h I ) 2 , where w I and h I represent the width and height of the intersection region, respectively. 4.5.3, Calculate the minimum area C of the bounding box and the real box by length and width, the formula is: C = (w C ) 2 + (h C ) 2 , where w C and h C are the width and height of the smallest area covering the predicted and real boxes, respectively, w C = max(L * , L) + max(R * , R) - 1, h C = max(T * , T) + max(B * , B) - 1. 4.5.

4. By minimizing c, the prediction box can move in 4 directions towards the ground truth box, eventually calculating SDIoU as where p represents a positive trade-off value that favors the overlapping area, and finally defines the SDIoU loss as loss = 1 - SDIoU. 5.The laparoscopic lesion target detection method based on VITDET class algorithm according to claim 1, wherein, In step 2, the specific steps of the histogram equalization algorithm are as follows: 2.1, scan each pixel of the original gray image in turn, and calculate the gray histogram of the image; 2.2, calculate the cumulative distribution function of the gray histogram; 2.3, get the mapping relationship between input and output according to the cumulative distribution function and the histogram equalization principle; 2.4, finally get the result according to the mapping relationship to transform the image. 6.The laparoscopic lesion target detection method based on VITDET class algorithm according to claim 1, wherein, In step 3, the distortion-free size modification method is: first scale the original image by a certain proportion, and then add a gray bar to the part that does not meet the size requirement.

Images