dMIL-Transformer: Multiple Instance Learning Via Integrating Morphological and Spatial Information for Lymph Node...

Area of Science:

• Computational pathology
• Medical image analysis
• Cancer diagnostics

Background:

• Automated classification of lymph node metastasis (LNM) is crucial for cancer diagnosis and prognosis.
• Accurate LNM classification requires integrating both tumor morphology and spatial distribution, which is challenging for existing methods.

Purpose of the Study:

• To propose a novel two-stage dMIL-Transformer framework for improved LNM classification.
• To effectively integrate morphological and spatial information of tumor regions using multiple instance learning (MIL).

Main Methods:

• A double Max-Min MIL (dMIL) strategy was developed to select critical positive instances from histopathology images.
• A Transformer-based MIL aggregator was designed to integrate instance-level information using self-attention mechanisms.
• The framework learns bag-level representations for predicting LNM category with enhanced interpretability.

Main Results:

• The dMIL strategy establishes a superior decision boundary for instance selection.
• The dMIL-Transformer framework achieved performance improvements of 1.79%-7.50% over state-of-the-art methods across three LNM datasets.
• The proposed method demonstrates effective handling of complex LNM classification tasks with strong visualization capabilities.

Conclusions:

• The dMIL-Transformer framework offers a robust and interpretable solution for automated LNM classification.
• Integrating morphological and spatial features via a two-stage MIL approach significantly enhances classification performance.
• This framework holds promise for advancing computational pathology in cancer diagnostics.