BiChemoCLAM: a weakly supervised multimodal framework for chemotherapy response prediction.

Area of Science:

• Computational biology
• Artificial intelligence in oncology
• Medical image analysis

Background:

• Chemotherapy is vital for cancer treatment but carries risks, necessitating accurate response prediction.
• Current weakly supervised learning methods struggle to integrate whole slide images and molecular data for chemotherapy response prediction, especially in small-sample scenarios.
• Effective integration of multimodal data is needed to enhance predictive power.

Purpose of the Study:

• To develop a novel multimodal deep learning framework, BiChemoCLAM, for interpretable and data-efficient chemotherapy response prediction.
• To effectively combine attention-driven multiple instance learning with multimodal compact bilinear pooling.
• To improve the accuracy of chemotherapy response prediction using integrated imaging and molecular data.

Main Methods:

• Developed BiChemoCLAM, a bimodal chemotherapy response multi-instance learning framework.
• Utilized attention-driven multiple instance learning to process whole slide images.
• Integrated gene expression data using multimodal compact bilinear pooling.
• Validated the framework on ovarian, colorectal, and bladder cancer datasets.

Main Results:

• BiChemoCLAM achieved Area Under Curve (AUC) scores of 80.91% for ovarian serous cystadenocarcinoma.
• The framework demonstrated AUCs of 71.68% for colorectal adenocarcinoma and 75.80% for bladder urothelial carcinoma.
• Experimental results confirm the effectiveness of BiChemoCLAM in predicting chemotherapy response.

Conclusions:

• BiChemoCLAM represents an effective multimodal deep learning approach for chemotherapy response prediction.
• The framework shows promise in addressing challenges of integrating imaging and molecular data in small-sample cancer datasets.
• This model offers a more interpretable and data-efficient solution for personalized cancer treatment strategies.