Deep learning based methods for constructing models for the identification of pulmonary mycosis and uses of the constructed models for identifying subjects having pulmonary mycosis

Abstract

This document discloses a method for establishing a model to identify pulmonary fungal infections (e.g., pulmonary aspergillosis). The method includes acquiring a raw chest X-ray image; preprocessing the raw image to generate a preprocessed image; generating a reconstructed normal chest X-ray image from the preprocessed image using a deep learning model; subtracting the reconstructed normal chest X-ray image from the corresponding image to generate a first heatmap, where positive values ​​on the first heatmap represent abnormal areas; masking areas on the first heatmap other than the lung lobes, leaving a second heatmap containing only the lung lobes; overlaying the second heatmap containing only the lung lobes onto the corresponding image to generate an image with labeled abnormal areas; feeding the image with labeled abnormal areas into a pre-trained DINO model to extract multiple features; and training a classifier using the extracted features to distinguish between normal, pulmonary fungal infections (PM), or other abnormalities (OA). This document also discloses a method for identifying individuals suffering from pulmonary fungal infections using the established model.

Claims

1. A method for establishing a model to identify pulmonary fungal infections, comprising: (1) providing an original chest X-ray image; (2) preprocessing the original image from step (1) to generate a preprocessed image; (3) reconstructing the preprocessed image from step (2) into a reconstructed normal chest X-ray image using a deep learning model; (4) subtracting the reconstructed normal chest X-ray image from step (3) from the preprocessed image from step (2) to generate a first heat map, wherein positive values ​​on the first heat map represent abnormal regions; (5) masking regions other than the lung lobes on the first heat map from step (4), leaving a second heat map containing only the lung lobe region; (6) superimposing the second heat map containing only the lung lobe region from step (5) onto the preprocessed image from step (2) to generate an image with labeled abnormal regions; (7) inputting the image with labeled abnormal regions from step (6) into a pre-trained DINO model to extract multiple features; and (8) The classifier is trained based on the features extracted in step (7) so that it can distinguish the original chest X-ray images from step (1) into normal, pulmonary fungal infection (PM) or other abnormal (OA) images, thereby establishing the model.

2. The method as described in claim 1, wherein the preprocessing in step (2) includes performing any of the following processing on the original image: inverting, adjusting contrast, removing noise, enhancing edges, removing borders, rotating, scaling, or any combination thereof.

3. The method as described in claim 1, wherein the deep learning model in step (3) is a convolutional neural network (CNN), a deep belief network (DBN), a deep Q-learning network (DQN), a diffusion model, a generative adversarial network (GAN), a graph neural network (GNN), a recurrent neural network (RNN), a variational autoencoder (VAE), a long short-term memory network (LSTM), or a hybrid model.

4. The method as described in claim 1, wherein the masking in step (5) is performed by the following steps: (5-1) cutting the reconstructed normal chest X-ray image of step (3) using an image segmentation model, and; (5-2) using the cutting result to mask the area on the first thermal image of step (4) other than the lung lobe, leaving a second thermal image containing only the lung lobe area.

5. The method as described in claim 4, wherein the image segmentation model is a U-Net model, an FCN model, a DeepLab model, or a SegNet model.

6. The method as described in Request 1, wherein the pre-trained DINO model in step (7) is a ViT model, a ResNet model, an EfficientNet model, or a CNN model.

7. The method as described in claim 6, wherein the classifier in step (8) can perform binary or ternary classification.

8. The method as described in claim 7, wherein the binary classification is capable of distinguishing whether the original image of the chest X-ray is normal or PM.

9. The method as described in claim 8, wherein the binary classification is performed by generating a probability (P) of normal or PM based on the following equation for prediction and classification: where, Each category i has a score Zi, which is used to calculate a softmax function to convert the original category scores into a probability (P) that sums to 1, where the category i is normal, M, or OA, and its score Zi is Z normal, ZPM, or ZOA.

10. The method as described in claim 7, wherein the ternary classification can distinguish the original image of the chest X-ray as normal, PM, or OA.

11. The method as described in claim 10, wherein the ternary classification is used to predict and classify by generating a probability (P) of normal, PM, and OA using the following equation: and each category i has a score Zi, which is used to calculate a softmax function to convert the original classification scores into a probability (P) that sums to 1, wherein the category i is normal, PM, or OA, and its score Zi is Znormal, ZPM, or ZOA.

12. The method as described in claim 1, wherein the pulmonary mycosis is pulmonary aspergillosis (PA).

13. A method for identifying whether an individual has pulmonary mycosis, comprising: (1) obtaining a chest X-ray image of the individual; (2) inputting the chest X-ray image obtained in step (1) into a model established by the method described in any one of claims 1-11, and generating an identification result of normal, pulmonary mycosis or other abnormality (OA).

Images