PCG-CAM: Enhanced class activation map using principal components of gradients and its applications in brain MRI

Abstract

Brain tumors are among the most prevalent and deadly diseases worldwide, making early diagnosis critical. However, existing automated brain tumor diagnostic methods often lack interpretability, and the high cost of labeled data limits their effectiveness. Class activation mapping (CAM) provides visual explanations and object localization for convolutional neural networks (CNNs) by highlighting regions of interest corresponding to specific classes. However, existing approaches tend to focus solely on discriminative regions and often contain excessive noise. In this paper, we propose a simpler and more efficient method called PCG-CAM, which provides visual explanations for brain tumor diagnosis and generates fine-grained pseudo-labels. PCG-CAM extracts the principal components of gradients and uses their absolute values as weights for the feature maps, thereby better reflecting the importance of each feature map while preserving more object features. We evaluated the saliency maps generated by PCG-CAM on weakly-supervised brain tumor segmentation and assessed their generalizability in object localization tasks. Specifically, our method achieves 47.42% mIoU in weakly-supervised brain tumor segmentation, outperforming other methods by nearly 10% on average. The results on brain MRI and natural images demonstrate that our method effectively localizes target positions and provides robust explanations for model decisions.