HAMIL: Hierarchical Attention Multi-Instance Learning for Label-Free Colorectal Cancer Typing

Area of Science:

• Oncology
• Biomedical Imaging
• Machine Learning

Background:

• Colorectal cancer (CRC) is a leading gastrointestinal malignancy requiring advanced diagnostic tools.
• Current pathological imaging for CRC diagnosis is time-consuming and requires expert annotation.
• Analyzing the tumor microenvironment is crucial for understanding CRC progression.

Purpose of the Study:

• To develop a label-free method for colorectal cancer typing using Hierarchical Attention Multi-Instance Learning (HAMIL).
• To integrate optical time-stretch (OTS) imaging with microfluidic cell focusing for high-throughput cell image acquisition.
• To construct a high-throughput CRC typing dataset for method validation.

Main Methods:

• Development of a high-throughput cell image acquisition system using optical time-stretch (OTS) imaging and microfluidic cell focusing.
• Construction of a CRC typing dataset comprising 363,931 cell images from 10 clinical samples.
• Implementation of HAMIL, featuring instance attention for single-cell analysis and bag attention for population-level characteristics.

Main Results:

• The HAMIL method achieved an 86.30% F1 score in CRC typing.
• HAMIL outperformed eight other advanced Multiple Instance Learning (MIL) methods.
• The model effectively captured tumor heterogeneity and microenvironment characteristics.

Conclusions:

• HAMIL provides an effective, label-free approach for clinical CRC typing.
• The integration of OTS imaging and HAMIL enables efficient analysis of cellular populations.
• This study establishes a new pathway for high-throughput analysis in gastrointestinal oncology.