Interpretability-Driven Sample Selection Using Self Supervised Learning for Disease Classification and Segmentation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Artificial Intelligence

Background:

• Effective sample selection is crucial for optimizing supervised learning in medical image analysis.
• Minimizing expert interactions, such as label querying in active learning, is essential for efficiency.

Purpose of the Study:

• To propose a novel sample selection methodology, Interpretability-Driven Sample Selection (IDEAL), for medical image analysis.
• To leverage deep features from interpretability saliency maps for identifying informative samples.

Main Methods:

• A novel self-supervised learning approach was developed to train a classifier for identifying the most informative samples.
• Three informativeness determination methods were analyzed: observational, radiomics-based, and self-supervised.
• The IDEAL approach was evaluated in active learning setups for lung disease classification and histopathology image segmentation.

Main Results:

• The proposed self-supervised approach within IDEAL demonstrated superior performance in selecting informative samples compared to baseline methods.
• IDEAL achieved state-of-the-art performance using fewer samples in both lung disease classification and histopathology segmentation tasks.
• The study confirmed the potential of using interpretability information for sample selection in active learning.

Conclusions:

• Interpretability-Driven Sample Selection (IDEAL), particularly its self-supervised component, offers a powerful strategy for enhancing active learning in medical image analysis.
• This method significantly improves efficiency by reducing the number of required expert labels while achieving high diagnostic accuracy.
• IDEAL represents a promising advancement for data-efficient AI in healthcare, applicable to diverse medical imaging tasks.