CN-117237269-B - Lung CT anomaly detection method based on multi-scale clipping and self-supervision reconstruction

Abstract

The invention relates to a lung CT anomaly detection method based on multi-scale clipping and self-supervision reconstruction. A self-supervised enhancement strategy CropMixPaste for lung CT scan images is first set to generate glass-like anomalies to encourage the model to detect local irregularities in the lung CT scan images, then a Simple Masking Attention Prediction Block (SMAPB) is inserted into the convolutional network, masking information is predicted by the self-supervised reconstruction module to refine local features, and finally an anomaly detector is constructed using representations learned in the self-supervised proxy task. Experimental results on a true lung CT scan dataset confirm the effectiveness and superiority of the method of the invention.

Inventors

• LI ZUOYONG
• LI WEI
• FAN HAOYI
• ZENG KUN
• LIU WEIXIA

Assignees

• 闽江学院

Dates

Publication Date

20260505

Application Date

20230214

Claims (6)

1. 1. A lung CT anomaly detection method based on multi-scale clipping and self-supervision reconstruction is characterized by firstly setting a self-supervision enhancement strategy CropMixPaste suitable for a lung CT scanning image to generate anomalies similar to ground glass shadows and encourage a model to detect local irregularities of the lung CT scanning image, then inserting a simple masking attention prediction block SMAPB into a convolution network to perform self-supervision learning to predict masking information to refine local features, finally constructing an anomaly detector by using a representation obtained by self-supervision learning, For a given set of normal training samples The method aims to distinguish a test sample as normal or abnormal, and the problem of abnormal detection is defined as follows: For test sets Each sample is associated with Dimension feature correlation The goal being to learn a scoring function To be based on threshold value For the sample Classification is carried out: Wherein the method comprises the steps of Representation of 0 Represents a normal class, and 1 represents an abnormal class; The CropMixPaste basic flow is as follows: (1) Obtaining a plurality of clipping blocks with different sizes from a normal image; (2) Blending the cropping block using Mixup to form a new blended image, wherein each blending operation is preceded by an intermediate enhancement; (3) Randomly blurring the blended image and randomly resizing the image, wherein a glass shadow-like anomaly is generated by the random blurring; (4) Pasting the image generated in the step (3) back to any position of the original image; the insertion of the simple masking attention prediction block SMAPB into the convolutional network, i.e., the introduction of the self-supervising reconstruction block into the self-supervising depth encoder The simple masked attention prediction block SMAPB consists of a masked convolution layer, also called a masked convolution filter, whose central region is masked, and a parametric attention free block, which ensures that the learned information is not too simplistic and that key features can be refined, to minimize reconstruction errors between the masked input and output, whose central region is masked, and whose sub-kernels at four corners are convolved to obtain an activation map to predict the center of the mask; the construction of anomaly detectors using self-supervised learned representations, i.e., from encoders via CropMixPaste, SMAPB Learned representations are used to construct an out-of-distribution detector for calculating anomaly scores Then, an anomaly score is obtained by testing the mahalanobis distance between the gaussian distribution of the test sample and the normal sample, as follows: Wherein the method comprises the steps of The average value is represented by a value of, Representing a Gaussian distribution In (1), wherein Is the dimension size, so the Markov distance matrix of the anomaly map can be calculated by utilizing the method And finally, determining the anomaly score of the whole image by using the highest score.
2. 2. The pulmonary CT anomaly detection method based on multi-scale clipping and self-supervised reconstruction of claim 1, wherein CropMixPaste is specifically implemented as follows: 1) Multi-scale cutting Given slave normal data Selected normal image of (a) Cutting block Expressed as: Wherein, the The cutting operation is represented by a cutting operation, Representing a resizing operation and then defining an inclusion Sets of individual clipping operations Wherein the upper limit and the lower limit of the clipping ratio are Dividing them into each clipping operation, capturing fine details by small clipping ratio operation, and covering more pixels by large clipping ratio to realize multi-scale information acquisition, thereby obtaining a group of numbers Is a different size cut block: 2) MixUp mixing operation To more naturally fuse different sized crop blocks to obtain multi-scale information, from Two cutting blocks are randomly selected 、 Mixing is performed by using MixUp mixing operation, and meanwhile, channel replacement is performed on the images with smaller mixing weights, which is defined as follows: Wherein the method comprises the steps of Representing channel alignment, mixing weights From the slave Is sampled at random in the middle of the period, Setting to 0.3, and repeating the above operation to mix the images With randomly selected cut blocks Mixing, mixing by repeated mixing operations The clipping blocks in the image are fused into a final image ; 3) Blurring and pasting To synthesize the image as much as possible With true image alignment, random blurring is used to simulate abnormal areas like the ground glass shadows GGO of the lungs, furthermore this operation can blur the edges of the blocks and facilitate continuity of pasting, then randomly blurring the image Is randomly resized and then pasted back to any position of the original image to form the final composite anomaly image ; 4) Training target for self-supervision learning By utilizing the normal sample and the abnormal sample generated by CropMixPatse, a deep neural network can be trained to learn useful characteristic representation so as to improve generalization capability of real abnormality and better detect CT scanning abnormality, therefore, the training target of self-supervision learning is defined as: Wherein the method comprises the steps of Is a group consisting of CropMixPaste a, In the case of a deep neural network, Representing cross entropy loss.
3. 3. The pulmonary CT anomaly detection method based on multi-scale clipping and self-supervised reconstruction of claim 1, wherein the masking convolution filter is implemented as follows: First, the masking region of the receptive field center is set to Wherein Representing the number of channels to Extend the distance to each direction for the center Child core capable of learning Disposed at four corners, thus, will create a size of Wherein ; Let the input tensor be Wherein Is the number of channels that are to be formed, And Height and width, respectively, and then using a designed masking convolution kernel for each pixel to obtain the entire tensor Specifically, the module only carries out convolution operation at the position of the subkernel of the receptive field, then adds the results of the four convolution operations, namely the masked result The method comprises the steps of performing convolution filtering on each channel to predict masking information from all channels, adding masking information around the input to ensure that the spatial dimensions of the input and the output are the same and each pixel is subjected to masking filtering Zero padding of individual pixels and setting the step size to 1, and finally activating the output tensor by means of a ReLU.
4. 4. A pulmonary CT anomaly detection method based on multi-scale clipping and self-supervised reconstruction as recited in claim 3, wherein the parametric-free attention module is configured to process the masked convolution filter output, i.e., using an attention module SimAM with full 3D weights, the following minimum energy equation is defined taking into account the energy function defined for each neuron: Wherein the method comprises the steps of , , Representing the super-parameters of energy, the lower the energy, the neuron The greater the distinction from the surrounding neurons, the greater the importance, and therefore, the importance of the neurons by The features are obtained and enhanced and optimized by using Sigmoid, and the definition is as follows: Wherein the method comprises the steps of All are put together Grouping is done across channels and spatial dimensions.
5. 5. The method of claim 1, wherein SMAPB has a separate reconstruction loss function to calculate the error between each masked region information and its reconstructed location information by a mask filter of the receptive field for an input tensor The mean square error MSE used is as follows: Wherein the method comprises the steps of The process of SMAPB is shown in the description, Representing the final output; the loss function of the method is thus as follows: Wherein the method comprises the steps of Is a hyper-parameter representing SMAPB loss of importance.
6. 6. The method for pulmonary CT anomaly detection based on a multi-scale cropping and self-monitoring reconstruction module of claim 1, wherein the method employs a technique that utilizes GradCAM visual interpretation to provide a visual anomaly heat map to guide the localization of anomaly defects.

Description

Lung CT anomaly detection method based on multi-scale clipping and self-supervision reconstruction Technical Field The invention belongs to the technical field of image processing, and particularly relates to a lung CT anomaly detection method based on multi-scale clipping and self-supervision reconstruction. Background Computed Tomography (CT) of the lungs is an effective imaging method that helps clinicians to quickly identify areas of lung infection. Studies have demonstrated that Computed Tomography (CT) has become one of the important tools to aid in diagnosis and treatment of viral pneumonia. Wherein, the chest CT image of the patient is highlighted by a ground glass shadow (GGO), and in severe cases, the chest CT image will be in a crazy paving shape. The lung is a CT abnormal sign which can occur in various diseases of the lung and is divided into two major categories of limitation and diffusion. Therefore, the development of the computer diagnosis auxiliary system to locate the abnormal region assists in finding the abnormal region which may be missed due to excessive work or inattention of the doctor, is helpful to improve the diagnosis efficiency and quality of patients with viral pneumonia, and prevents the wide and rapid spread of diseases. Currently, various deep learning methods have made significant progress in CT scan image analysis. However, supervised deep learning is limited by the availability of data sets and computing resources. In most cases, the number of lung CT images is limited and the resolution is low. At the same time, time consuming and expensive manual annotation of CT scan images requires knowledge of medical professionals. Therefore, an unsupervised anomaly detection method without additional tags has been the focus of research. Unlike typical supervised classification problems, pulmonary CT anomaly detection faces two challenges. First, anomalies are agnostic, making it difficult to obtain extensive anomaly data (e.g., rare lesions). Second, the differences between normal and abnormal images in medical images are finer than those in natural data sets, requiring that the model have sensitivity to identify both minor differences and salient features. In recent years, scholars at home and abroad have made a series of researches on the problem of abnormality detection of medical images. Three general categories of methods, reconstruction-based methods, feature similarity-based methods, and self-supervised learning-based methods, are possible. The reconstruction-based method trains an encoder network (AE), a generative model (GAN), or a normalized Flow (Flow) to reconstruct similar samples. Since the model only obtains the characteristic distribution of normal samples, the difference between the generated or reconstructed samples and the input is an abnormal region. In addition, the feature similarity-based method obtains the anomaly score by calculating the similarity between the feature vectors of the normal training sample and the test sample, and the common similarity is provided with Gaussian distribution, a representative repository, a nearest neighbor algorithm and the like. Meanwhile, some methods also adjust the characteristics of the training sample through the technologies of characteristic coupling, early stopping, distillation and the like so as to obtain more distinguishable normal characteristics. Recently, methods based on self-supervised learning have demonstrated excellent performance. Such methods learn better feature representations by setting proxy tasks, such as predictive geometric transformations, predictive rotations, and predictive data enhancements. Thanks to the targeted setting of the agent task, the method is widely applied to auxiliary diagnosis technologies such as brain magnetic resonance, lung CT, abdomen CT, chest X-ray, eye OCT and the like, and has certain application value. Thus, anomaly detection of lung CT aims to distinguish the anomaly image from the lung CT scan image and locate the ground glass shadow (GGO) anomaly region, as shown in fig. 1. Disclosure of Invention The invention aims to provide a lung CT anomaly detection method based on multi-scale cutting and self-supervised reconstruction, which aims to improve the anomaly detection performance of lung CT scan images. In order to achieve the above purpose, the technical scheme of the invention is that a lung CT abnormality detection method based on multi-scale clipping and self-supervision reconstruction is characterized in that firstly, a self-supervision enhancement strategy CropMixPaste suitable for a lung CT scanning image is set to generate abnormal similar to ground glass shadows, a model is encouraged to detect local irregularities of the lung CT scanning image, then a simple masking attention prediction block SMAPB is inserted into a convolution network to conduct self-supervision learning to predict masking information to refine local features, and finally, an abnormalit