Active and Inactive Tuberculosis Classification Using Convolutional Neural Networks with MLP-Mixer.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Infectious Disease Diagnostics

Background:

• Early detection of tuberculosis (TB) is crucial for effective treatment and preventing reactivation.
• Identifying inactive TB forms, like latent or healed TB, is essential for proactive management.
• Conventional diagnostic methods may have limitations in rapid and accurate differentiation of TB states.

Purpose of the Study:

• To develop and evaluate a deep-learning model for binary classification of active versus inactive tuberculosis cases.
• To assess the model's performance in distinguishing between different forms of tuberculosis for improved early detection.
• To provide a tool that supports clinical decision-making in TB screening and management.

Main Methods:

• Development of a deep-learning binary classification model using an EfficientNet backbone and MLP-Mixer classification head.
• Fine-tuning the model on a dataset from Cheonan Soonchunhyang Hospital.
• Application of transfer learning with weights pre-trained on the JFT-300M dataset using the Noisy Student training method.

Main Results:

• The deep learning model achieved high performance on the test set.
• Achieved an accuracy of 96.3%, sensitivity of 95.9%, and specificity of 96.6%.
• Demonstrated competitive results compared to conventional models in distinguishing active from inactive TB.

Conclusions:

• The developed deep learning model shows significant potential for supporting clinical decision-making in tuberculosis diagnosis.
• The model can streamline early screening workflows, particularly for latent tuberculosis.
• This AI-driven approach offers a promising advancement in the early identification and management of tuberculosis.