Tumor Bud Classification in Colorectal Cancer Using Attention-Based Deep Multiple Instance Learning and...

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology

Background:

• Tumor budding (TB) in colorectal cancer (CRC) is a critical prognostic indicator linked to metastasis.
• Manual TB assessment is subjective and challenging, particularly at high magnifications, leading to diagnostic inconsistencies.
• Automated methods are needed to improve the accuracy and reproducibility of TB identification.

Purpose of the Study:

• To develop and evaluate an automated deep learning system for accurate tumor budding classification in colorectal cancer.
• To compare the performance of various foundation models for feature extraction in TB detection.
• To enhance the prognostic assessment of colorectal cancer through improved TB identification.

Main Methods:

• A deep learning model was trained using weakly supervised learning on whole slide images (WSIs) from the tumor invasive front.
• Multiple foundation models were assessed for feature extraction, with performance compared.
• Attention heatmaps from attention-based multi-instance learning (ABMIL) were analyzed for interpretability and alignment with TBs.

Main Results:

• Phikon-v2 achieved the highest average AUC (0.984) and recall (0.947) in cross-validation.
• On the external test set, Phikon-v2 demonstrated superior performance with an AUC of 0.979 and precision of 0.980.
• The model's attention mechanisms provided interpretable insights into TB identification.

Conclusions:

• The proposed automated system significantly enhances the accuracy of tumor budding assessment in colorectal cancer.
• This deep learning approach offers a promising tool for improving prognostic evaluation in CRC.
• The technique has potential applications for TB assessment in other cancer types.