BPMambaMIL: A bio-inspired prototype-guided multiple instance learning for oncotype DX risk assessment in...

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Genomic assay alternatives

Background:

• Hormone receptor-positive/human epidermal growth factor receptor 2-negative breast cancer is common.
• Genomic assays like Oncotype DX guide treatment but are costly and inaccessible.
• Whole slide images (WSIs) are large and computationally challenging to process.

Purpose of the Study:

• To develop a cost-effective method for predicting Oncotype DX risk using H&E-stained pathology images.
• To introduce a novel weakly supervised learning framework, BPMambaMIL, for this prediction task.
• To improve breast cancer recurrence risk assessment and accessibility.

Main Methods:

• Developed BPMambaMIL, a bio-inspired prototype-guided model integrating Mamba and prototypical guidance.
• Utilized weakly supervised learning on H&E-stained whole slide images (WSIs).
• Evaluated model performance on in-house and public breast cancer datasets.

Main Results:

• BPMambaMIL achieved an AUC of 0.839 on an in-house dataset, outperforming the baseline MambaMIL by 5.61%.
• Demonstrated robust performance in identifying high-risk score ranges (accuracy: 0.714).
• Showcased generalizability on public datasets for Oncotype DX scores and binary tumor classification.

Conclusions:

• BPMambaMIL accurately predicts breast cancer recurrence risk from pathology images.
• The model offers a cost-effective alternative to expensive genomic assays.
• This approach can improve clinical decision-making and patient outcomes.