Explainable unsupervised TNM category differentiation in PET images with deep texture analysis

We propose an integrated approach to TNM category classification in PET images that combines ROI detection via a U-Net with unsupervised manifold learning. First, the U-Net identifies candidate regions of interest (ROIs) for primary tumors, nodes, and metastases. A patch-based convolutional neural network (CNN), pretrained on tumor data, then extracts metabolic deep texture features, and Gradient-weighted Class Activation Mapping (Grad-CAM) highlights salient regions within each patch. These Grad-CAM heatmaps are subsequently embedded into a low-dimensional space using Uniform Manifold Approximation and Projection (UMAP), and clustered into T, N, and M classes with k-means. This unsupervised pipeline eliminates the need for extensive node and metastasis annotations, as only tumor patches are used for the initial supervised training. Our method achieves 89.5% accuracy and a 93.1% F1 score in distinguishing T, N, and M categories, indicating robust performance. By combining the interpretability of Grad-CAM with data-driven manifold learning, this approach shows promise for enhancing esophageal cancer staging and aiding more personalized treatment planning.