Fungus drug sensitive plate image processing and classifying method

Abstract

The invention provides a fungal drug sensitive plate image processing and classifying method, which comprises the following steps of: image cutting: acquiring a fungal drug sensitive plate image, and uniformly cutting the fungal drug sensitive plate image into a plurality of sub-images; classification prediction: each subgraph adopts a pre-trained ResNet-34 model to perform image classification, the pre-trained ResNet-34 model can return a classification label of each subgraph, and each classification label represents a category; sorting labels: sorting the sorting labels of the sub-images into a matrix according to the positions of the sub-images in the fungal drug sensitive plate image to obtain a sorting label matrix; 01 matrix construction: the 01 matrix is constructed according to the arranged classification label matrix, and each element in the 01 matrix represents the classification result of the holes in the corresponding fungal drug sensitive plate. The method has high efficiency and accuracy when being used for processing the fungal drug sensitive plate image, the average time consumption for processing one image is 0.2 s, and the liquid plate hole positive and negative classification accuracy is higher than 95%.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to a fungus drug-sensitive plate image processing and classification method. Background Art

[0002] In recent years, due to the increase in the number of patients with reduced body resistance such as transplantation, tumors, and immunodeficiency, as well as the widespread use of broad-spectrum antimicrobial drugs, the body's flora has been imbalanced, and clinical fungal infections have been on the rise. Common clinical fungal infections include Candida, Aspergillus, Cryptococcus, etc., which can be confined to one organ or present as multi-organ disseminated infections. Severe invasive fungal infections can invade various organs and systems of the human body such as the heart, lungs, blood, and gastrointestinal tract, endangering the patient's life. Therefore, antifungal infection diagnosis and treatment have become one of the research hotspots, especially invasive fungal infections have attracted increasing attention. Candidiasis and new cryptococcosis are the two most common fungal diseases. Candidiasis is mainly an acute, subacute or chronic infection caused by Candida albicans, which often invades the skin and mucous membranes, and can also cause visceral or systemic infections. The clinical symptoms are complex and vary in severity. When patients have symptoms of Candida infection, medical staff will often use drugs such as amphotericin B, fluconazole, and caspofungin to treat patients. New cryptococcosis can invade humans and animals. It is usually an exogenous infection or an endogenous infection. For humans, it is usually conditional. When the human immune function declines, the new cryptococcus can spread throughout the body, mainly invading the central nervous system, causing meningitis, encephalitis, brain granulomas, etc. In addition, it can invade bones, muscles, lymph nodes, skin and mucous membranes to cause chronic inflammation and abscesses. The therapeutic drugs widely used by medical staff include amphotericin B, 5-flucytosine, fluconazole, itraconazole, voriconazole, etc. With the continuous emergence of new antifungal drugs and clinical fungal resistance, empirical antifungal treatment is becoming more and more difficult. There is an urgent need for fungal drug sensitivity tests to guide clinical drug use, prognosis judgment, and evaluate the efficacy of new drugs. There are many methods for antifungal drug sensitivity tests, among which the micro-liquid dilution method has the advantages of being simple and easy to perform and sensitive, and is welcomed by most clinical workers. However, manual interpretation of the endpoint of fungal susceptibility tests is time-consuming and laborious, requiring high professional skills of personnel, and the interpretation standards greatly affect the accuracy of the test results. There is an urgent need for a standardized intelligent interpretation system to make the test results of different laboratories comparable.

[0003] With the maturity of neural network applications, the efficiency of using them for classification and identification has been greatly improved. Therefore, how to apply neural networks to fungal drug sensitivity plate image processing and classification has become an important issue at present. Traditional image processing methods usually require a lot of manual feature extraction and rule definition, which is difficult to adapt to the processing of large-scale data. Neural networks have the advantages of high efficiency, automation and intelligence, high precision and accuracy in image processing. Neural networks can automatically learn features from images. Through a large amount of training data and optimization algorithms, they can recognize and distinguish complex patterns and features in images. Feature extraction through multiple levels helps to understand the image more comprehensively, making the image analysis results more accurate and reliable. Applying artificial intelligence to the interpretation of fungal drug sensitivity plate results greatly improves the accuracy of the results, while saving time and improving detection efficiency. Summary of the invention

[0004] Purpose of the invention: In view of the above-mentioned prior art, a method for image processing and classification of fungal drug sensitivity plates is proposed;

[0005] Technical solution: A method for processing and classifying fungal drug sensitivity plates, comprising the following steps:

[0006] 1) Image cutting: Obtain the fungal drug sensitivity plate image and evenly cut the fungal drug sensitivity plate image into several sub-images;

[0007] 2) Classification prediction: For each sub-image in step 1), a pre-trained ResNet-34 model is used for image classification. The pre-trained ResNet-34 model returns the classification label of each sub-image, where each classification label represents a category;

[0008] 3) Classification label arrangement: according to the position of the sub-image in the fungal drug sensitivity plate image in step 1), the classification label of each sub-image is arranged into a matrix to obtain a classification label matrix;

[0009] 4) 01 matrix construction: According to the arranged classification label matrix, a 01 matrix is ​​constructed, where each element in the 01 matrix represents the classification result of the corresponding well on the fungal drug sensitivity plate, where 0 in the 01 matrix represents negative and 1 represents positive.

[0010] Preferably, in step 1), the several subgraphs cut evenly correspond to the holes on the fungal drug sensitivity plate, and satisfy:

[0011] W sub =W total / 12

[0012] H sub =H total / 8

[0013] Where W total and Htotal are the width and height of the fungal drug sensitivity plate image, W sub and H sub are the width and height of the sub-image respectively.

[0014] Preferably, in step 2),

[0015] The ResNet-34 model includes multiple residual blocks, each of which includes two branches: the main branch and the skip branch. The output of the residual block is the element-wise addition of the main branch output and the skip branch output:

[0016] x l+1 =x l +F(x l ; Θ l )

[0017] where x l+1 represents the output of the l+1th convolutional layer in the ResNet-34 model, x l represents the output of the lth convolutional layer in the ResNet-34 model, Θ l represents the parameters of the lth convolutional layer, and F(.) is a function;

[0018] The output x of the deep network of the ResNet-34 model L It is expressed as:

[0019]

[0020] Among them, x i represents the i-th residual block, Θ i represents the parameters of the i-th residual block, L=34 is the number of convolutional layers;

[0021] The ResNet-34 model was supervised pre-trained on the ImageNet dataset and fine-tuned on the fungal drug sensitivity plate image dataset. During fine-tuning, the number of output categories of the last fully connected layer of ResNet-34 was adjusted to 2, and only the parameters of this layer of the model were trained, and the remaining model parameters were frozen. The cross entropy loss function L was used for both training and fine-tuning. CE :

[0022]

[0023] Where -log(p i ) is the probability output by the ResNet-34 model, y i is the true label, and C is the parameter.

[0024] Beneficial effects: It is efficient and accurate in processing fungal susceptibility plate images. It takes an average of 0.2 seconds to process one image, and the accuracy of liquid plate well positive and negative classification is higher than 95%. DETAILED DESCRIPTION

[0025] The present invention will be further explained below.

[0026] A fungus drug sensitivity plate image processing and classification method comprises the following steps:

[0027] 1) Image cutting: Obtain the fungal drug sensitivity plate image and evenly cut the fungal drug sensitivity plate image into several sub-images;

[0028] 2) Classification prediction: For each sub-image in step 1), a pre-trained ResNet-34 model is used for image classification. The pre-trained ResNet-34 model returns the classification label of each sub-image, where each classification label represents a category;

[0029] 3) Classification label arrangement: according to the position of the sub-image in the fungal drug sensitivity plate image in step 1), the classification label of each sub-image is arranged into a matrix to obtain a classification label matrix;

[0030] 4) 01 matrix construction: According to the arranged classification label matrix, a 01 matrix is ​​constructed, where each element in the 01 matrix represents the classification result of the corresponding well on the fungal drug sensitivity plate, where 0 in the 01 matrix represents negative and 1 represents positive.

[0031] In step 1), the uniformly cut subgraphs correspond to the holes on the fungal drug sensitivity plate and satisfy:

[0032] W sub =W total / 12

[0033] H sub =H total / 8

[0034] Where W total and H total are the width and height of the fungal drug sensitivity plate image, W sub and H sub are the width and height of the sub-image respectively.

[0035] In step 2),

[0036] The ResNet-34 model includes multiple residual blocks, each of which includes two branches: the main branch and the skip branch. The output of the residual block is the element-wise addition of the main branch output and the skip branch output:

[0037] x l+1 =x l +F(x l ; Θ l )

[0038] where x l+1 represents the output of the l+1th convolutional layer in the ResNet-34 model, x l represents the output of the lth convolutional layer in the ResNet-34 model, Θ l represents the parameters of the lth convolutional layer, and F(.) is a function;

[0039] The output x of the deep network of the ResNet-34 model L It is expressed as:

[0040]

[0041] Among them, x i represents the i-th residual block, Θ i represents the parameters of the i-th residual block, L=34 is the number of convolutional layers;

[0042] The ResNet-34 model was supervised pre-trained on the ImageNet dataset and fine-tuned on the fungal drug sensitivity plate image dataset. During fine-tuning, the number of output categories of the last fully connected layer of ResNet-34 was adjusted to 2, and only the parameters of this layer of the model were trained, and the remaining model parameters were frozen. The cross entropy loss function L was used for both training and fine-tuning. CE :

[0043]

[0044] Where -log(p i ) is the probability output by the ResNet-34 model, y i is the true label and C is the parameter.

[0045] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for processing and classifying fungal drug sensitivity plates, characterized in that: The following steps are involved: 1) Image cutting: Obtain the fungal drug sensitivity plate image and evenly cut the fungal drug sensitivity plate image into several sub-images; 2) Classification prediction: For each sub-image in step 1), a pre-trained ResNet-34 model is used for image classification. The pre-trained ResNet-34 model returns the classification label of each sub-image, where each classification label represents a category; 3) Classification label arrangement: according to the position of the sub-image in the fungal drug sensitivity plate image in step 1), the classification label of each sub-image is arranged into a matrix to obtain a classification label matrix; 4) 01 matrix construction: According to the arranged classification label matrix, a 01 matrix is ​​constructed, where each element in the 01 matrix represents the classification result of the corresponding well on the fungal drug sensitivity plate, where 0 in the 01 matrix represents negative and 1 represents positive.

2. A fungal drug-sensitive plate image processing and classification method as claimed in claim 1, characterized in that: In step 1), the several subgraphs cut evenly correspond to the holes on the fungal drug sensitivity plate and satisfy: IN sub =In total / 12 H sub =H total / 8 Where W total and H total are the width and height of the fungal drug sensitivity plate image, W sub and H sub are the width and height of the sub-image respectively.

3. A fungus drug-sensitive plate image processing and classification method as claimed in claim 2, characterized in that: In step 2), The ResNet-34 model includes multiple residual blocks, each of which includes two branches: the main branch and the skip branch. The output of the residual block is the element-wise addition of the main branch output and the skip branch output: x l+1 =x l +F(x l ;Θ l ) where x l+1 represents the output of the l+1th convolutional layer in the ResNet-34 model, x l represents the output of the lth convolutional layer in the ResNet-34 model, Θ l represents the parameters of the lth convolutional layer, and F(.) is a function; The output x of the deep network of the ResNet-34 model L It is expressed as: Among them, x i represents the i-th residual block, Θ i represents the parameters of the i-th residual block, L=34 is the number of convolutional layers; The ResNet-34 model was supervised pre-trained on the ImageNet dataset and fine-tuned on the fungal drug sensitivity plate image dataset. During fine-tuning, the number of output categories of the last fully connected layer of ResNet-34 was adjusted to 2, and only the parameters of this layer of the model were trained, and the remaining model parameters were frozen. The cross entropy loss function L was used for both training and fine-tuning. CE : Where -log(p i ) is the probability output by the ResNet-34 model, y i is the true label and C is the parameter.

Images