Weakly-Supervised Transfer Learning with Application in Precision Medicine.

Area of Science:

• Machine Learning
• Precision Medicine
• Medical Imaging

Background:

• Precision medicine requires personalized models for improved diagnosis and treatment.
• Limited labeled data per individual poses a challenge for personalized machine learning.
• Transfer Learning (TL) addresses data scarcity by leveraging information from similar patient groups (source domain).

Purpose of the Study:

• To introduce Weakly-Supervised Transfer Learning (WS-TL) to overcome limitations in existing TL algorithms.
• To address scenarios with minimal or no labeled data in the target domain.
• To integrate domain knowledge effectively into TL for personalized models.

Main Methods:

• Developed a novel mathematical framework for WS-TL using domain knowledge-inferred paired samples.
• Integrated labeled data from the source domain for knowledge transfer.
• Proposed an efficient active sampling strategy to select informative paired samples.

Main Results:

• WS-TL demonstrated superior accuracy compared to existing TL algorithms in a real-world brain cancer application.
• Personalized patient models for Tumor Cell Density (TCD) prediction were successfully developed.
• The method effectively handles target domains with few or no labeled samples.

Conclusions:

• WS-TL offers a robust solution for building accurate personalized models in precision medicine, even with limited individual data.
• The approach facilitates individually optimized treatment strategies through precise TCD mapping.
• This work advances the application of transfer learning in medical AI.