Developing an explainable diagnosis system utilizing deep learning model: a case study of spontaneous pneumothorax.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Pulmonary Disease Diagnostics

Background:

• Intelligent diagnostic systems lack interpretability, creating 'black box' issues that hinder misdiagnosis identification and treatment improvement.
• Lack of interpretability in medical AI poses risks of misdiagnosis and delayed treatment, impacting patient outcomes.
• Limited research exists on deep learning for spontaneous pneumothorax prediction, a condition affecting lung ventilation and venous return.

Purpose of the Study:

• To develop an interpretable medical image analysis system for automatic diagnosis.
• To enhance the accuracy and trustworthiness of diagnostic models through explainable AI.
• To address the need for interpretable deep learning in spontaneous pneumothorax detection.

Main Methods:

• Development of an integrated medical image analysis system.
• Implementation of an explainable deep learning model for image recognition and visualization.
• Focus on achieving an interpretable automatic diagnosis process.

Main Results:

• The system achieved 95.56% accuracy in classifying spontaneous pneumothorax.
• Identified the significance of blood vessel penetration defects in clinical judgment.
• Demonstrated the potential for improved model trustworthiness and reduced diagnostic uncertainty.

Conclusions:

• The explainable deep learning system offers accurate diagnosis of lung diseases, improving patient outcomes.
• Enhanced model interpretability leads to better utilization of medical resources.
• Future work can extend this system to diagnose other lung diseases, increasing generalizability.