Random features meet MIL: a deep GP approach to colorectal MSI prediction

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Colorectal cancer (CRC) poses a significant global health challenge, with early diagnosis being critical for patient survival.
• Accurate prediction of CRC from medical images is hindered by challenges like weak supervision, data heterogeneity, and large datasets.

Purpose of the Study:

• To develop a novel deep learning approach for colorectal cancer classification using weakly labeled whole-slide images.
• To improve the accuracy, robustness, and interpretability of automated CRC detection models.

Main Methods:

• Integration of deep Gaussian processes (DGP) with multi-instance learning (MIL) to handle bag-level labels.
• Utilizing a deep Gaussian process with random feature expansion (DGP-RF) for improved classification performance.
• Incorporating an attention-based aggregation mechanism to focus on critical regions within whole-slide images.

Main Results:

• The proposed model achieved a superior Area Under the Curve (AUC) of 0.895 on the TCGA-CRC dataset.
• Outperformed established models like ResNet (0.777), EfficientNet (0.791), and ShuffleNet (0.784).
• Demonstrated enhanced accuracy and robustness in colorectal cancer classification.

Conclusions:

• The novel DGP-MIL approach offers a significant advancement in automated colorectal cancer detection from medical images.
• The model's performance suggests potential for clinical deployment in improving early diagnosis and patient outcomes.