Towards Interpretable, Sequential Multiple Instance Learning: An Application to Clinical Imaging.

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Sequential data processing

Background:

• Interpreting sequential medical images with variable lengths and single labels is challenging.
• Traditional Multiple Instance Learning (MIL) methods often overlook the inherent sequence order in clinical imaging.

Purpose of the Study:

• Introduce the Sequential Multiple Instance Learning (SMIL) framework to address sequential medical image interpretation.
• Develop a model that integrates sequence order for improved diagnostic accuracy and efficiency.
• Introduce an interpretable uncertainty metric for enhanced model evaluation.

Main Methods:

• Developed a bidirectional Transformer architecture (BiSMIL) tailored for sequential medical image data.
• Implemented a novel training procedure to optimize both early and final prediction accuracies.
• Introduced SMILU, a new uncertainty metric for evaluating model performance on challenging instances.

Main Results:

• BiSMIL achieved state-of-the-art final accuracy on three medical image datasets.
• Demonstrated superior early prediction accuracy, requiring 30-50% fewer images than existing models.
• The SMILU metric outperformed traditional metrics in identifying difficult cases.

Conclusions:

• SMIL framework effectively leverages sequential information in medical imaging.
• BiSMIL offers a balance between diagnostic accuracy and operational efficiency.
• SMILU provides a valuable tool for assessing model reliability in medical AI.