Using Multi-Scale Convolutional Neural Network Based on Multi-Instance Learning to Predict the Efficacy of...

Area of Science:

• Digital pathology
• Machine learning in oncology
• Rectal cancer research

Background:

• Radical total mesorectal excision is vital for locally advanced rectal cancer post-neoadjuvant chemoradiotherapy.
• Histopathological image analysis can predict treatment efficacy, guiding patient management.

Purpose of the Study:

• To develop and validate a novel pathological image analysis method for predicting neoadjuvant chemoradiotherapy efficacy in rectal cancer.
• To assist clinicians in rapid diagnosis and personalized treatment planning.

Main Methods:

• Utilized multi-instance learning with a gated attention normalization mechanism for faster training.
• Employed a bilinear attention multi-scale feature fusion mechanism to preserve context.
• Incorporated a weighted loss function to address class imbalance.

Main Results:

• Achieved AUC values of 0.9337 on the Camelyon16 dataset and 0.9091 on the MSKCC dataset.
• Demonstrated strong generalization performance on multiple datasets.
• Validated on a locally advanced rectal cancer dataset of 150 whole slide images.

Conclusions:

• The proposed method shows outstanding performance in predicting neoadjuvant chemoradiotherapy efficacy for rectal cancer.
• This technology can significantly aid clinicians in making informed treatment decisions.