A novel multiple-instance learning-based approach to computer-aided detection of tuberculosis on chest X-rays

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Detection

Background:

• Supervised learning for CAD systems requires extensive manual lesion annotation, which is time-consuming and difficult to obtain.
• This limitation hinders the development and adaptability of CAD systems in real-world medical applications.

Purpose of the Study:

• To investigate multiple-instance learning (MIL) as an alternative to supervised learning for tuberculosis detection using CAD systems.
• To propose an improved MIL algorithm addressing limitations of existing methods like miSVM.
• To evaluate the performance and adaptability of the MIL-based CAD system compared to a traditional supervised approach.

Main Methods:

• Applied MIL to a CAD system for tuberculosis detection, utilizing only case-level (normal/abnormal) labels for training.
• Developed an improved algorithm based on miSVM to overcome positive instance underestimation and costly iterations.
• Conducted experiments on three X-ray databases, comparing the MIL-based system with a supervised system using Area Under the ROC Curve (AUC).

Main Results:

• The MIL-based method achieved performance comparable to the supervised system when detailed annotations were available (AUC 0.86 vs. 0.88).
• In scenarios requiring retraining with limited data (case-level labels only), the MIL-based system demonstrated superior adaptability.
• After retraining, the MIL-based system significantly outperformed the supervised system on two databases (AUC 0.86 vs. 0.79 and 0.91 vs. 0.85, p=0.0002).

Conclusions:

• MIL provides a viable and adaptable alternative to supervised learning for CAD systems, especially when detailed annotations are scarce.
• The proposed improved MIL algorithm enhances performance and efficiency, making it suitable for real-world tuberculosis detection applications.
• The study highlights the practical advantages of MIL-based CAD systems in adapting to new data and varying conditions common in clinical practice.