A Deep Learning-Based Approach for Explainable Microsatellite Instability Detection in Gastrointestinal Malignancies.

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Microsatellite instability (MSI) is a crucial biomarker in gastrointestinal cancers, impacting diagnosis and treatment.
• Current MSI detection methods are expensive, slow, and require specialized labs.
• There is a need for efficient and accessible MSI detection techniques.

Purpose of the Study:

• To develop an explainable deep learning (DL) method for MSI detection from histopathological images.
• To evaluate the performance of various DL models, including CNNs and Vision Transformers.
• To enhance the trustworthiness of DL predictions through explainability and robustness analysis.

Main Methods:

• Utilized histopathological images for MSI detection.
• Implemented and compared several DL architectures: MobileNet, Inception, VGG16, VGG19, and Vision Transformer.
• Employed three Class Activation Mapping (CAM) techniques for model interpretability.
• Introduced robustness metrics to assess the consistency of highlighted regions across CAM methods.

Main Results:

• VGG16 and VGG19 models demonstrated superior performance, achieving accuracies of 0.926 and 0.917, respectively.
• CAM techniques confirmed that models consistently focused on relevant tissue regions.
• Robustness analysis indicated high agreement between different CAM methods, strengthening prediction reliability.

Conclusions:

• The proposed DL method accurately detects MSI in gastrointestinal cancers using histopathological images.
• The approach provides explainable predictions, increasing confidence in model outputs.
• This method has the potential to facilitate the adoption of DL in clinical practice for cancer diagnostics.