REACT-KD: Region-Aware Cross-modal Topological Knowledge Distillation for Interpretable Medical Image Classification

TL;DR

REACT-KD tackles the problem of reliable, interpretable tumor grading under heterogeneous clinical imaging by transferring supervision from high-fidelity multi-modal sources to a lightweight CT-based student. It introduces a dual-teacher framework (structure–function-aware PET/CT and dose-aware LDCT) and a region graph distillation (RGD) module that enforces topology-aware alignment via ROI-based graphs and a Gromov-Wasserstein loss. A shared CBAM3D attention mechanism and modality dropout during training enhance cross-modal fusion and robustness to missing inputs. On hepatocellular carcinoma grading, REACT-KD achieves an internal AUC of $93.5\%$ and maintains strong external performance ($AUC>0.76$) under varying dose degradation, while providing topology-guided interpretability of decisions.

Abstract

Reliable and interpretable tumor classification from clinical imaging remains a core challenge. The main difficulties arise from heterogeneous modality quality, limited annotations, and the absence of structured anatomical guidance. We present REACT-KD, a Region-Aware Cross-modal Topological Knowledge Distillation framework that transfers supervision from high-fidelity multi-modal sources into a lightweight CT-based student model. The framework employs a dual teacher design. One branch captures structure-function relationships through dual-tracer PET/CT, while the other models dose-aware features using synthetically degraded low-dose CT. These branches jointly guide the student model through two complementary objectives. The first achieves semantic alignment through logits distillation, and the second models anatomical topology through region graph distillation. A shared CBAM3D module ensures consistent attention across modalities. To improve reliability in deployment, REACT-KD introduces modality dropout during training, which enables robust inference under partial or noisy inputs. As a case study, we applied REACT-KD to hepatocellular carcinoma staging. The framework achieved an average AUC of 93.5\% on an internal PET/CT cohort and maintained 76.6\% to 81.5\% AUC across varying levels of dose degradation in external CT testing. Decision curve analysis further shows that REACT-KD consistently provides the highest net clinical benefit across all thresholds, confirming its value in real-world diagnostic practice. Code is available at: https://github.com/Kinetics-JOJO/REACT-KD

Figures (4)

• Figure 1: Overall pipeline of our distillation framework. (a) Dual-branch teacher model: a structure-function-aware (SF) path processes dual-tracer PET/CT images from the institutional cohort, and a dose-aware path handles simulated dose-varied CT scans derived from the LiTS17 dataset. (b) CT features from both paths are extracted using a shared SwinUNETR encoder, while PET features are separately encoded. (c) Feature fusion is performed via a shared CBAM3D module to enhance discriminative regions. (d) Region graphs are constructed based on liver and tumor masks for topological supervision. (e) A student network learns under modality dropout with guidance from logits-based and region graph distillation losses.
• Figure 2: Decision curve analysis (DCA, left) on internal validation and receiver operating characteristic (ROC, right) on external validation. The DCA curves indicate sustained net clinical benefit across decision thresholds, confirming the model’s practical reliability under degraded CT conditions, while the ROC curves demonstrate consistent discriminative capability across tumor grades.
• Figure 3: t-SNE projection of student logits on external test set. Each subfigure corresponds to one cross-validation fold. The model trained with logits-only supervision yields incomplete separation of tumor grades, with particularly blurred margins between moderate and other differentiation levels. These findings highlight the insufficiency of global prediction signals in capturing nuanced class boundaries.
• Figure 4: Topology-aware region graph visualization on external cases. Each row depicts one test subject from the external cohort. From left to right: original CT image, anatomical ROIs (liver in red, tumors in green), attention maps from the student trained without region graph distillation (RGD), and the student trained with RGD.