MICIL: Multiple-Instance Class-Incremental Learning for skin cancer whole slide images

Area of Science:

• Computer Science
• Medical Imaging
• Artificial Intelligence

Background:

• Artificial intelligence (AI) agents face catastrophic forgetting when trained sequentially, hindering applications like cancer prediction.
• Whole slide images (WSIs) are vital for cancer management, driving demand for automated analysis.
• Incremental learning (IL) aims to enable AI models to learn continuously without forgetting past information.

Purpose of the Study:

• To introduce a novel IL algorithm for analyzing gigapixel WSIs within a multiple instance learning (MIL) framework.
• To address the challenge of catastrophic forgetting in AI models applied to evolving datasets, specifically for cancer prediction.
• To enable incremental prediction of multiple skin cancer subtypes from WSIs in a class-incremental learning (class-IL) scenario.

Main Methods:

• Developed the Multiple Instance Class-Incremental Learning (MICIL) algorithm, combining MIL with class-IL for WSI analysis.
• Incorporated knowledge distillation, data rehearsal, and a novel embedding-level distillation to preserve latent space.
• Evaluated the algorithm's performance on class-IL metrics, focusing on balancing plasticity and stability.

Main Results:

• The MICIL algorithm effectively addresses catastrophic forgetting in WSI analysis.
• Demonstrated the ability to incrementally predict multiple skin cancer subtypes from WSIs.
• Successfully balanced IL-specific metrics, mitigating intransigence and forgetting.

Conclusions:

• The proposed MICIL algorithm represents a significant advancement in applying IL to WSI analysis for cancer prediction.
• This framework offers a robust solution for continuous learning in medical imaging, preserving knowledge while adapting to new data.
• The study highlights the potential of combining MIL and class-IL for scalable and effective AI in digital pathology.