REACT-KD[BIBM 2025 Regular]

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

REACT-KD[BIBM 2025 Onsite-viedo]

B1189-Chen_9.3MB.mp4

📄Preprint version

Our paper has been accepted by BIBM 2025 as a regular paper:

📌 Pipeline Overview

The current pipeline reflects our initial implementation. LoRA-based components are excluded due to space limitations. A LoRA-enabled student model is provided for reference, and future work will explore LoRA-C for robust CNN vs Transformer adaptation.

Citation

If this code or our framework is useful for your research, please consider citing our paper:

@misc{chen2025reactkdregionawarecrossmodaltopological,
      title={REACT-KD: Region-Aware Cross-modal Topological Knowledge Distillation for Interpretable Medical Image Classification}, 
      author={Hongzhao Chen and Hexiao Ding and Yufeng Jiang and Jing Lan and Ka Chun Li and Gerald W. Y. Cheng and Sam Ng and Chi Lai Ho and Jing Cai and Liang-ting Lin and Jung Sun Yoo},
      year={2025},
      eprint={2508.02104},
      archivePrefix={arXiv},
      primaryClass={eess.IV},
      url={https://arxiv.org/abs/2508.02104}, 
}

🧠 Teacher Model

🎯 Student Model

📚 Dataset Description

To enable multi-modal knowledge distillation and external validation in hepatocellular carcinoma (HCC) tumor grade classification, we utilized a diverse dataset composed of:

🏥 1. Hospital HCC Cohort (Private)

• Source: In house-HK SAR population
• Period: January 2004 – December 2024
• Patients: 97 HCC patients
• Imaging: 194 dual-tracer PET/CT scan pairs
  • PET tracers: ¹⁸F-FDG and ¹¹C-Acetate
  • Co-registered PET + CT volumes (512 × 512 × 148)
• Labels:
  • Tumor Pathology grade (Edmondson–Steiner classification: 3 classes(Well differentiated, Moderately differentiated, Poorly differentiated))
  • Clinical metadata: AFP, age, sex, hepatitis status

---

📊 2. LiTS17 Benchmark (Public)

• Source: LiTS 2017 Challenge
• Scans: 131 contrast-enhanced abdominal CT volumes
• Annotations: Liver and tumor segmentation masks
• Usage: Encoder-level pretraining and topological graph construction
• Resolution:
  • Slice range: 42–1026 slices
  • In-plane spacing: 0.56–1.0 mm

---

🧪 3. HCC-TACE-Seg Dataset (Public)

• Scans: 105 multiphasic contrast-enhanced CT scans
• Annotations: Expert segmentation + histopathological tumor grade (3 classes: Well differentiated, Moderately differentiated, Poorly differentiated)
• Context: Real-world TACE treatment under standard imaging protocols
• Usage: External validation only

---

🔄 Dataset Usage Summary

Dataset	Role	Modality	Grade Label	Notes
HKH PET/CT (priv)	Teacher training	PET + CT	✅ Yes	Main supervised distillation source
LiTS17	Teacher training	CT only	❌ No	Used for segmentation + graph prep
HCC-TACE-Seg	External validation	CT only	✅ Yes	No training, validation only