Search

CN-121998905-A - Chest radiography image abnormal region detection method and program product

CN121998905ACN 121998905 ACN121998905 ACN 121998905ACN-121998905-A

Abstract

The invention provides a chest radiography image abnormal region detection method and a program product, wherein a FAHG module is newly added on the basis of YOLOv and used for carrying out frequency domain enhancement and noise suppression on a feature image output to a detection head, a FAHG module is used for carrying out Fourier transform on the feature image, then respectively carrying out low-pass, band-pass and high-pass filtering to decompose three sub-band components of low frequency, medium frequency and high frequency, carrying out enhancement processing on the three sub-band components, wherein the low-frequency and medium-frequency sub-band components are information interaction and feature expression for enhancing a cross channel, the high-frequency sub-band component is noise suppression, and carrying out weighted fusion after respectively carrying out Fourier inverse transformation on the three sub-band components after enhancement. Compared with other existing YOLO series models, the chest radiography image abnormal region detection method has the advantages of being high in detection precision and generalization capability, capable of remarkably improving the detection precision without obviously increasing the calculated amount, and well balanced between efficiency and precision.

Inventors

  • Liu Yanglijuan
  • YAN XIANGYANG
  • GUAN LI
  • GUO HAO
  • LOU LEI
  • CAO RUI

Assignees

  • 南京工业职业技术大学

Dates

Publication Date
20260508
Application Date
20251223

Claims (10)

  1. 1. A chest radiography image abnormal region detection method is characterized in that the following image detection model is adopted: A FAHG module is newly added on the basis of YOLOv < 11 >, and the FAHG module is used for carrying out frequency domain enhancement and noise suppression on the feature map output to the detection head; the FAHG module includes the following processing: S1, performing Fourier transform on the feature map; S2, respectively performing low-pass, band-pass and high-pass filtering on the converted characteristics to decompose three subband components of low frequency, medium frequency and high frequency; s3, carrying out enhancement processing on three sub-band components, wherein the low-frequency and medium-frequency sub-band components are the information interaction and characteristic expression of the enhanced cross-channel, and the high-frequency sub-band components are the noise suppression; s4, respectively performing inverse Fourier transform on the three enhanced sub-band components of low frequency, intermediate frequency and high frequency, and then performing weighted fusion.
  2. 2. The method for detecting abnormal regions of chest radiography according to claim 1, wherein in S3, the low-frequency band component and the medium-frequency band component are respectively enhanced by a channel aggregation module, and the processing procedure of the channel aggregation module is as follows: in the formula, And The input and output of the modules are respectively, As a parameter of the weight-bearing element, The representation GELU activates the function, Representing a1 x 1 convolution.
  3. 3. The method for detecting abnormal regions of chest radiography according to claim 1, wherein in S3, the high-frequency subband components are enhanced by a squeeze-excitation module, and the process of the squeeze-excitation module is as follows: in the formula, And The input and output of the modules are respectively, Representing the Sigmoid activation function, Is a full-connection layer, and is formed by the following steps, Representing the function of the ReLU activation, Representing the pooling process.
  4. 4. The method for detecting abnormal regions of a chest radiography image according to claim 1, wherein S1 is a fast Fourier transform.
  5. 5. The method for detecting abnormal regions of chest radiography images according to claim 1, wherein in the image detection model, the original C3k2 module is modified as follows: Under the condition that the C3k2 module is in a False state, the Bottleneck module in the original C3k2 module is replaced by a MSPLC module; under the state that the C3k2 module is True, replacing the C3k module in the original C3k2 module with a C3_ MSPLC module; In the MSPLC module, for input quantity, firstly, sequentially carrying out batch normalization and two-layer convolution treatment, then respectively carrying out parallel treatment through three cavity convolutions with different convolution kernel sizes, and then carrying out fusion through 1 multiplied by 1 convolution after the obtained results are spliced; The C3_ MSPLC module is divided into two parallel paths for input quantity, wherein one path is processed by the continuous multilayer MSPLC module after being subjected to 1X 1 convolution, the other path is processed by the 1X 1 convolution, and the fusion is carried out by the 1X 1 convolution after the last two paths of results are spliced.
  6. 6. The method for detecting abnormal regions of chest radiography according to claim 5, wherein the MSPLC module processes: in the formula, And Respectively representing the input quantity and the output quantity of the module, The feature stitching is represented and is performed, 、 、 Respectively represent the convolution of holes with convolution kernel sizes of 3 multiplied by 3, 5 multiplied by 5 and 7 multiplied by 7, A5 x 5 convolution is indicated and is shown, Indicating batch normalization.
  7. 7. The method for detecting abnormal regions of a chest radiography image according to claim 6, wherein the expansion rates of the cavity convolutions are all 3.
  8. 8. The method for detecting abnormal regions of chest radiography images according to claim 1, wherein the image detection model adopts the following feature fusion module to replace feature stitching operation in an original neck network: Firstly, splicing according to the channel dimension; The spliced feature map is multiplied element by the weight vector.
  9. 9. The method of claim 8, wherein the weight vector is normalized by Softmax.
  10. 10. A computer program product comprising a computer program which, when executed by a processor, implements the chest radiography image anomaly region detection method according to any one of claims 1 to 9.

Description

Chest radiography image abnormal region detection method and program product Technical Field The invention belongs to the technical field of medical image processing and computer-aided diagnosis, and particularly relates to a chest radiography image abnormal region detection method and a program product. Background Chest radiography is the most widely used screening means for chest diseases in clinical practice. Currently, radiologists have heavy manual film reading work and are easy to miss or misdiagnose due to fatigue and experience differences. In this regard, computer vision based automated detection techniques help alleviate this problem. However, the application of the computer vision technology in the field of chest film focus detection has the special challenges that 1) focus scale difference is huge, and the focus scale difference is coexistent from tiny nodules to large-scale real changes, 2) the contrast of focuses and normal tissues is often low, the boundary is fuzzy, 3) the image background is complex, the overlapping interference of anatomical structures such as ribs, blood vessels and the like is serious, and 4) focuses of different categories (such as nodules, exudation and fibrosis) can appear in the image simultaneously and overlap each other. The current mainstream solutions are mostly based on Convolutional Neural Networks (CNNs), especially YOLO series single-stage detectors. While these methods have advantages in terms of speed, their feature extraction and fusion mechanisms are typically designed for generic objects and do not adequately accommodate the above-described characteristics of chest radiography images. Specifically, the existing method is limited to spatial domain operation, lacks utilization of image frequency domain information to separate structure and noise, is fixed or single in receptive field when multi-scale features are fused, is difficult to flexibly capture focus with extremely large span, and is usually simple to splice or add when features of different network levels are fused, so that key features are diluted or noise is amplified. Therefore, the detection precision and recall rate of the existing YOLO neural network on small-size and low-contrast lesions in chest radiography images still are insufficient. Disclosure of Invention Aiming at the defects existing in the prior art, the invention provides a chest radiography image abnormal region detection method and a program product, and the detection precision of small-size and low-contrast lesions in chest radiography images is improved based on YOLOv improvement. The present invention achieves the above technical object by the following technical means. A chest radiography image abnormal region detection method adopts the following image detection model: A FAHG module is newly added on the basis of YOLOv < 11 >, and the FAHG module is used for carrying out frequency domain enhancement and noise suppression on the feature map output to the detection head; the FAHG module includes the following processing: S1, performing Fourier transform on the feature map; S2, respectively performing low-pass, band-pass and high-pass filtering on the converted characteristics to decompose three subband components of low frequency, medium frequency and high frequency; s3, carrying out enhancement processing on three sub-band components, wherein the low-frequency and medium-frequency sub-band components are the information interaction and characteristic expression of the enhanced cross-channel, and the high-frequency sub-band components are the noise suppression; s4, respectively performing inverse Fourier transform on the three enhanced sub-band components of low frequency, intermediate frequency and high frequency, and then performing weighted fusion. Further, in the step S3, the low frequency subband components and the medium frequency subband components are respectively enhanced by a channel aggregation module, where a processing procedure of the channel aggregation module is as follows: in the formula, AndThe input and output of the modules are respectively,As a parameter of the weight-bearing element,The representation GELU activates the function,Representing a1 x 1 convolution. Further, in the step S3, the high-frequency subband components are enhanced by the extrusion-excitation module, where the extrusion-excitation module processes: in the formula, AndThe input and output of the modules are respectively,Representing the Sigmoid activation function,Is a full-connection layer, and is formed by the following steps,Representing the function of the ReLU activation,Representing the pooling process. Further, in S1, a fast fourier transform is performed. Further, in the image detection model, the original C3k2 module is modified as follows: Under the condition that the C3k2 module is in a False state, the Bottleneck module in the original C3k2 module is replaced by a MSPLC module; under the state that the C3k2 module is True