CN-121983332-A - Artificial intelligent recognition system and application thereof in fungus morphological diagnosis

Abstract

The invention relates to an artificial intelligent recognition system and application thereof in fungus morphological diagnosis, and relates to the technical field of medical inspection and artificial intelligent intersection. The artificial intelligent recognition system deeply fuses an immunofluorescence technology and artificial intelligence, can be used for clearly distinguishing dead/living bacteria, and performs image recognition, segmentation and feature extraction by using the artificial intelligent system, so that comprehensive diagnosis is performed by combining clinical data.

Inventors

• CHEN HAISHENG
• ZHAO YANGXI

Assignees

• 佛山大学

Dates

Publication Date

20260505

Application Date

20251231

Claims (10)

1. 1. An artificial intelligence recognition system is characterized by comprising an adaptive segmentation module, a preprocessing module, an AI recognition module and a multi-layer perceptron; The self-adaptive segmentation module is used for inputting a multi-mode image and outputting a multi-scale local patch, the image preprocessing module is used for inputting the multi-scale local patch and outputting a normalized multi-scale local patch, the AI recognition module is used for inputting the normalized multi-scale local patch and outputting target characteristics, and the multi-layer sensor is used for inputting the target characteristics and real data for analysis.
2. 2. The artificial intelligence recognition system of claim 1, wherein the adaptive segmentation module comprises a U-net++ adaptive segmentation network, the U-net++ adaptive segmentation network being optimized from a U-net++ architecture, the optimization being achieved by a weighted mixing loss function L total .
3. 3. The artificial intelligence recognition system of claim 2, wherein the weighted mixing loss function L total is as follows: L total =α*L Dice (P fungi ,G fungi )+β*LCE(P all ,G all ) Wherein, alpha and beta are adjusting weight super parameters, L Dice is a Dice loss function, P fungi is predicted target pixel probability distribution, G fungi is a target real labeling pixel set, P all is predicted pixel probability distribution of all categories, G all is a predicted real labeling pixel set of all categories, and LCE is a cross entropy loss function.
4. 4. The artificial intelligence recognition system of claim 1, wherein the image preprocessing module normalizes the multi-scale local plaque by Z-score normalization and contrast-limited adaptive histogram equalization to obtain a normalized multi-scale local plaque.
5. 5. The artificial intelligence recognition system of claim 1, wherein the AI recognition module comprises a deep learning framework with a convolved block attention module, the deep learning framework being R-CNN or ResNet-50.
6. 6. A method for automated diagnosis of fungal morphology based on an artificial intelligence recognition system according to any one of claims 1 to 5, comprising the steps of: s1, acquiring images, namely acquiring multi-mode images of secretion samples to be tested and clinical intervention data of sample providers; S2, automatic feature reasoning, namely inputting a multi-mode image into the artificial intelligent recognition system, actively recognizing and separating secretion impurities through the self-adaptive segmentation module, and extracting target features by utilizing the AI recognition module; And S3, outputting a comprehensive result, wherein the artificial intelligent recognition system outputs a morphological diagnosis report of fungi in the secretion sample to be detected.
7. 7. The automated diagnostic method of fungus morphology according to claim 6, wherein the morphological diagnostic report of the fungus comprises the survival status, classification of species and morphological quantitative analysis result of the fungus.
8. 8. The method for constructing an artificial intelligence recognition system according to any one of claims 1 to 5, comprising the steps of: Sample preparation, namely collecting clinical samples, marking by a fluorescent staining method, collecting fluorescent images under multiple fluorescent channels, enhancing data, and forming a multi-mode image database; Inputting the multi-mode image into a self-adaptive segmentation module to generate a global plaque, randomly extracting a plurality of local plaques with different scales, and obtaining a multi-scale local plaque; inputting the multi-scale local plaque into an image preprocessing module for normalization processing to obtain a normalized multi-scale local plaque; inputting the normalized multi-scale local plaque into an AI recognition module, and extracting target features; And (3) manual interaction labeling, namely inputting the target characteristics and the real data into a multi-layer perceptron, and carrying out weight iteration and model training through a supervised learning algorithm until a system loss function converges, so as to complete the construction of the artificial intelligent recognition system.
9. 9. The method according to claim 8, wherein in the image preprocessing step, normalization processing is performed by using Z-score normalization and contrast-limited adaptive histogram equalization.
10. 10. A computer device comprising a memory storing a computer program and a processor that when executing the computer program implements the automated fungal morphology diagnostic method of claim 6.

Description

Artificial intelligent recognition system and application thereof in fungus morphological diagnosis Technical Field The invention relates to the technical field of medical inspection and artificial intelligence intersection, in particular to an artificial intelligence identification system and application thereof in fungus morphological diagnosis. Background With the popularization of immunosuppressive treatment and organ transplantation and aging of population, the incidence and mortality rate of invasive fungal infection continue to rise, and the disease progress is rapid, which forms a serious threat to human health. Clinically traditional fungus diagnosis methods, such as morphological manual microscopic examination, culture method, serological detection and the like, have obvious limitations. For example, the culture method has long time consumption and possibly delays the optimal treatment time, and the manual microscopic examination is highly dependent on the professional experience of the inspector, so that the subjectivity is strong, the result consistency is poor (the variation coefficient CV in the group can reach 15-20%), and the time for culturing a qualified morphological expert is 5-10 years, so that the cost is high. In recent years, artificial Intelligence (AI) technology, particularly a deep learning algorithm, has great potential in the field of medical image analysis, can rapidly and accurately extract features from massive image data, and provides a new technical path for solving the diagnosis dilemma. At present, a scheme combining modern science and technology with morphological detection has become a research hot spot, for example, a mold morphological identification technology based on a neural network, namely, research on morphological identification of clinically common mold by using a residual neural network (Residual Neural Network) is available. According to the scheme, morphological characteristics such as hyphae, spores and the like are automatically extracted through a deep learning model, so that automatic classification of moulds is realized. The method is based on the principle that a neural network is trained by using a large number of marked mould microscopic images, so that the network learns the morphological differences of different moulds, and then a new unknown sample is predicted and identified by using a trained model. However, existing AI techniques for identifying fungi suffer from (1) inadequate processing power for complex clinical specimens. The existing AI morphological analysis models are mostly trained and tested under ideal conditions with clean background and few interferents. And the clinical true samples (such as sputum and tissue fluid) often contain a large amount of impurities such as blood, mucus and cell debris, and the impurities can generate background fluorescence or morphological interference, so that the recognition accuracy of the existing algorithm is obviously reduced. (2) the model generalization ability is limited. The existing model is usually trained on a data set of a single medical center, and when applied to other hospitals, the performance of the model is greatly reduced due to the differences of microscope models, staining schemes, operation procedures and the like, namely the generalization capability is poor. This limits the wide spread and application of AI diagnostic techniques. (3) The application scene is single and is not deeply fused with the immunofluorescence technology. Although AI is used in the study of mold morphology recognition, AI systems directed specifically to immunofluorescence images, in particular, capable of distinguishing the states of live/dead fungal cells and performing accurate quantitative analysis, are not yet mature. Disclosure of Invention In order to solve the problems, the invention provides an artificial intelligent recognition system, which is deeply fused with immunofluorescence technology and artificial intelligence, can be used for clearly distinguishing dead/living bacteria, and performs image recognition, segmentation and feature extraction by using the artificial intelligent recognition system, so as to perform comprehensive diagnosis by combining clinical data. The invention provides an artificial intelligent recognition system, which comprises a self-adaptive segmentation module, a preprocessing module, an AI recognition module and a multi-layer perceptron, wherein the self-adaptive segmentation module is used for preprocessing the AI recognition module; The self-adaptive segmentation module is used for inputting a multi-mode image and outputting a multi-scale local patch, the image preprocessing module is used for inputting the multi-scale local patch and outputting a normalized multi-scale local patch, the AI recognition module is used for inputting the normalized multi-scale local patch and outputting target characteristics, and the multi-layer sensor is used for inputting the targe