Brain tumor classification model guided by class activation mapping

Brain tumors present significant risks to human health, with various types differing in treatment and prognosis. Accurate diagnosis is essential for improving survival rates. Multimodal Magnetic Resonance Imaging plays a key role in tumor identification but faces challenges such as similar intensity distributions across modalities, tumor morphological variability, and indistinct boundaries. Existing classification methods often lack interpretability, which is crucial for clinical decision-making.

This study proposes a brain tumor classification model that integrates Class Activation Mapping techniques to enhance interpretability. The model generates stable class activation maps through end-to-end training, which act as weak supervision signals for tumor localization. Key components include a Saliency Learning Module for improved attention weight generation, a Sample Selection Module for better sample differentiation, and a Balanced Perception Loss function to guide model training.

The proposed model achieved classification accuracies ranging from 96% to 99%, averaging 97.41% across ten-fold cross-validation. Precision, recall, and F1 score averaged 97.53%, 97.66%, and 97.58%, respectively. These results demonstrate robust performance in differentiating various brain tumor types. The model's use of class activation maps enhances its interpretability, enabling visualization of the decision-making process and pinpointing tumor regions. Additionally, the model performed better than other methods in comparative experiments.

Our brain tumor classification model improves accuracy and interpretability. It provides a significant advancement in brain tumor diagnosis by accurately localizing tumor regions and facilitating clinicians' understanding of model predictions.