Brain tumor classification model guided by class activation mapping.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Brain tumors pose significant health risks, necessitating accurate diagnosis for improved patient outcomes.
• Multimodal Magnetic Resonance Imaging (MRI) is crucial for tumor identification but faces challenges like similar intensity distributions and indistinct boundaries.
• Current classification methods often lack the interpretability required for clinical decision-making.

Purpose of the Study:

• To develop an accurate and interpretable brain tumor classification model.
• To enhance clinical decision-making through visualized diagnostic processes.
• To address limitations of existing methods in handling complex MRI data.

Main Methods:

• Proposed a novel brain tumor classification model integrating Class Activation Mapping (CAM) for enhanced interpretability.
• Employed end-to-end training to generate stable CAMs for tumor localization, acting as weak supervision.
• Incorporated a Saliency Learning Module, Sample Selection Module, and Balanced Perception Loss function.

Main Results:

• Achieved high classification accuracies (96%-99%, average 97.41%) across ten-fold cross-validation.
• Demonstrated robust performance with average precision (97.53%), recall (97.66%), and F1 score (97.58%).
• CAMs provided enhanced interpretability, visualizing decision-making and pinpointing tumor regions, outperforming other methods.

Conclusions:

• The developed model significantly improves brain tumor classification accuracy and interpretability.
• Accurate tumor localization and visualized predictions facilitate better clinical understanding and diagnosis.
• Represents a substantial advancement in AI-driven brain tumor diagnosis.