MIST: Multi-instance selective transformer for histopathological subtype prediction.

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning for healthcare

Background:

• Accurate histopathological subtype prediction is vital for cancer diagnosis and understanding the tumor microenvironment.
• Challenges include instance-level discrimination, high intra-class variance, and heterogeneous feature distributions in histopathological images.

Purpose of the Study:

• To develop a novel framework for accurate histopathological subtype prediction using fine-grained representation learning.
• To address the challenges of instance discrimination and feature heterogeneity in histopathological image analysis.

Main Methods:

• Proposed the Multi-Instance Selective Transformer (MIST) framework, integrating Multi-Instance Learning (MIL) and Vision Transformer (ViT).
• Introduced a selective self-attention mechanism to identify informative instances.
• Developed modules for instance-to-instance and instance-to-bag interaction modeling to learn discriminative representations.

Main Results:

• The MIST framework achieved state-of-the-art performance on five clinical benchmarks.
• Demonstrated accurate histopathological subtype prediction with an accuracy of 0.936.
• Showcased the framework's effectiveness in handling fine-grained medical image analysis.

Conclusions:

• The MIST framework offers a powerful approach for accurate histopathological subtype prediction.
• Highlights the potential of MIST in clinical applications for fine-grained medical image analysis.
• The proposed selective self-attention and interaction modeling effectively address key challenges in histopathological image analysis.