AI-Driven Thoracic X-ray Diagnostics: Transformative Transfer Learning for Clinical Validation in Pulmonary Radiography

Abstract

Our research evaluates advanced artificial (AI) methodologies to enhance diagnostic accuracy in pulmonary radiography. Utilizing DenseNet121 and ResNet50, we analyzed 108,948 chest X-ray images from 32,717 patients and DenseNet121 achieved an area under the curve (AUC) of 94% in identifying the conditions of pneumothorax and oedema. The model's performance surpassed that of expert radiologists, though further improvements are necessary for diagnosing complex conditions such as emphysema, effusion, and hernia. Clinical validation integrating Latent Dirichlet Allocation (LDA) and Named Entity Recognition (NER) demonstrated the potential of natural language processing (NLP) in clinical workflows. The NER system achieved a precision of 92% and a recall of 88%. Sentiment analysis using DistilBERT provided a nuanced understanding of clinical notes, which is essential for refining diagnostic decisions. XGBoost and SHapley Additive exPlanations (SHAP) enhanced feature extraction and model interpretability. Local Interpretable Model-agnostic Explanations (LIME) and occlusion sensitivity analysis further enriched transparency, enabling healthcare providers to trust AI predictions. These AI techniques reduced processing times by 60% and annotation errors by 75%, setting a new benchmark for efficiency in thoracic diagnostics. The research explored the transformative potential of AI in medical imaging, advancing traditional diagnostics and accelerating medical evaluations in clinical settings.

Keywords: artificial intelligence; deep learning; diagnostic accuracy; medical imaging; model interpretability; pulmonary radiography.

Figure 22

LIME Analysis for Image Data (a) on original image to LIME explanation.

Figure 23

LIME Analysis for Image Data (b) on original image to LIME explanation.

Figure 24

LIME Analysis for Image Data (c) on original image to LIME explanation.

Figure 25

LIME Analysis for Image Data (d) on original image to LIME explanation.

Figure 26

LIME Analysis for Image Data (e) on original image to LIME explanation.

Figure 27

LIME Analysis for Image Data (f) on original image to LIME explanation.

Figure 28

Occlusion Sensitivity Map Analysis on labeled (a–f) image data. This figure demonstrates the occlusion sensitivity map analysis applied to labeled image data. By systematically occluding different parts of the image and observing the changes in the model’s predictions, this analysis helps identify the most crucial regions that influence the model’s decision-making process.

Figure 1

Data Assembly Process. The diagram illustrates the comprehensive steps involved in assembling the dataset, including data collection from hospital radiology departments, anonymization by removing patient identifiers to ensure ethical compliance, initial labeling using automated natural language processing (NLP) techniques, manual verification by radiologists, and rigorous quality control processes before finalizing the dataset.

Figure 2

Visualization of RGB data using Imshow, with RGB values normalized to the range [0, 1] for float representation. This figure illustrates how Imshow processes and displays RGB color data accurately.

Figure 3

Detection of COVID-19 in X-ray images using convolutional neural networks (CNNs). This figure illustrates the process and results of using CNNs to identify COVID-19 related anomalies in chest X-rays, highlighting the areas of the lungs affected by the virus.

Figure 4

Image pre-processing in Keras. This figure demonstrates the steps involved in pre-processing images using the Keras library, including resizing, normalization, and augmentation techniques to prepare the images for training in a neural network.

Figure 5

Distribution of classes for the training dataset and the relationship between values and classes. This figure illustrates the frequency of each class within the training dataset, along with a comparison of various values against these classes to provide insights into the dataset composition and balance.

Figure 6

Distribution of pixel intensity of an image. This figure displays the histogram of pixel intensities, illustrating the frequency of each intensity level across the image. It provides insights into the image’s contrast, brightness, and overall tonal distribution.

Figure 7

Implementation of weight loss in neural networks. This figure demonstrates the process of incorporating weight loss functions during the training phase of neural networks to prevent overfitting and improve generalization.

Figure 8

Training a neural network using DenseNet121. This figure illustrates the process of training a model with the DenseNet121 architecture, highlighting key steps such as data input, model configuration, and training iterations. DenseNet121 is known for its dense connectivity between layers, which can improve gradient flow and feature reuse.

Figure 9

Visualizing learning with Grad-CAM. This figure demonstrates the use of Grad-CAM to visualize which regions of an image contribute most to the neural network’s prediction. By highlighting important areas, Grad-CAM helped in understanding and interpreting the decision-making process of the model.

Figure 10

Image segmentation probability visualization. This figure illustrates the probability maps generated during the image segmentation process, showing the likelihood of each pixel belonging to different segments. It provided insights into the model’s confidence and accuracy in distinguishing various regions within the image.

Figure 11

Feature Pyramid Network (FPN) architecture with ResNet module. This figure illustrates the integration of the FPN architecture with the ResNet module, demonstrating how feature maps are extracted at multiple scales and combined to improve object detection performance. The FPN enhances the model’s ability to detect objects of varying sizes by leveraging the hierarchical feature representation of ResNet.

Figure 12

Feature extraction map on original image to feature map visualisaiton.

Figure 13

Chest Xray potential area mark on original image to potential lung area.

Figure 14

Class imbalance in the dataset. This figure highlights the distribution of classes within the dataset, illustrating the prevalence of class imbalance. Such imbalance can affect the performance of machine learning models by biasing predictions towards the majority class.

Figure 15

Class frequencies in the dataset. This figure shows the frequency of each class within the dataset, illustrating the distribution and relative abundance of different classes. Understanding class frequencies is crucial for addressing class imbalance and ensuring fair and accurate model training.

Figure 16

AUC values for the CheXNeXt model and radiologists on the dataset. This figure compares the AUC values for the CheXNeXt model and human radiologists, highlighting the performance of the deep learning model in diagnosing medical conditions from chest X-ray images relative to expert radiologists.

Figure 17

Evaluation of the DenseNet121 model results using the ROC curve. This figure displays the ROC curve for the DenseNet121 model, illustrating the model’s performance in distinguishing between classes by plotting the true positive rate against the false positive rate at various threshold settings. The ROC curve helps assess the model’s diagnostic accuracy.

Figure 18

Evaluation of the ResNet50 model results using the ROC curve. This figure displays the ROC curve for the ResNet50 model, illustrating the model’s performance in distinguishing between classes by plotting the true positive rate against the false positive rate at various threshold settings. The ROC curve helps assess the model’s diagnostic accuracy.

Figure 19

Training and validation accuracy and loss for the VGG19 model. This figure presents the training and validation accuracy, as well as the loss metrics, over multiple epochs during the training of the VGG19 model. It highlights the model’s learning progress and performance, showing how well the model generalizes to unseen data and identifying potential overfitting.

Figure 20

Results of sentiment analysis. This figure illustrates the outcomes of a sentiment analysis performed on a dataset, showcasing the distribution of positive, negative, and neutral sentiments. It highlights the model’s ability to classify text data based on emotional tone, providing insights into the overall sentiment trends within the dataset.

Figure 21

LIME: This figure illustrates the use of LIME to explain the predictions of a machine learning model. By highlighting the most influential features, LIME provides insights into how the model makes decisions, thereby enhancing interpretability and trust in the model’s outputs.