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

Colorectal cancer (CRC) is one of the leading gastrointestinal malignancies, underscoring the need for an in-depth analysis of the cellular within the tumor microenvironment. While pathological imaging remains the gold standard for cancer diagnosis, it requires extensive annotation time and expert knowledge. Therefore, we propose hierarchical attention multi-instance learning (HAMIL) for label-free CRC typing. Specifically, we integrate optical time-stretch (OTS) imaging technology with microfluidic cell focusing to develop a high-throughput cell image acquisition system, enabling efficient collection of CRC cell images. We measure 10 clinical samples, including 5 from normal samples and 5 from cancerous samples, resulting in a total of 363,931 cell images to construct a high-throughput CRC typing dataset. Based on the clinical CRC typing dataset, our proposed HAMIL utilizes an instance attention layer to extract instance attention weights from individual single-cell instances, allowing for fine-grained modeling of tumor heterogeneity and the tumor microenvironment. Building upon these instance attention weights, the bag attention layer integrates bag-level feature representations, capturing the overall characteristics of the high-throughput cellular population on a global scale. The experimental results show that HAMIL exceeds eight advanced MIL methods and reaches an 86.30% F1 score, which is expected to provide an effective new pathway for clinical CRC typing.