LymphAware

🧬 LymphAware: Domain-Aware Bias Disruption for Reliable Lymphoma Cancer AI Diagnosis

Teerapong Panboonyuen
College of Computing, Khon Kaen University
🎓 Supported by the Talent Scholarship for Exceptional Ability

Peer-Reviewed & Accepted at IEEE Access (February 2026) 🎉
DOI: 10.1109/ACCESS.2026.3667575

LymphAware Architecture

🚀 Overview

LymphAware is a domain-aware bias disruption framework designed to improve the reliability, robustness, and clinical relevance of AI systems for lymphoma histopathology diagnosis.

Modern medical AI models often achieve high accuracy by exploiting non-biological shortcuts — such as stain color, scanner signatures, or slide artifacts — instead of true pathological morphology. While effective in-domain, these shortcuts lead to fragile performance under cross-center variability, which is unacceptable for clinical deployment.

LymphAware explicitly addresses this challenge by separating morphology-relevant signals from shortcut-driven acquisition factors, enabling models to “think more like pathologists.” 🧠🔬

✨ Key Innovations

🔹 Tri-Path Morphology Purification Architecture

Morphology-centric feature encoder
Shortcut identification & suppression branch
Cross-domain stability alignment stream

🔹 Artifact-Shift Counterfactual Training

Simulated staining and scanner perturbations
Exposure of latent shortcut dependencies
Acquisition-invariant representation learning

🔹 Domain-Aware Robustness Without Explicit Labels

Works under realistic multi-source settings
No assumption of verified institutional separation

📊 Qualitative Results — Shortcut Suppression

Qualitative Results

Models trained without LymphAware rely heavily on stain tone, background artifacts, and acquisition noise. With LymphAware, attention shifts toward diagnostically meaningful lymphoid morphology.

📈 Cross-Center Performance

Performance Tables

Across five independent medical centers:

✅ Higher AUC ✅ Lower false positive rates ✅ Reduced variance across backbones ✅ Stronger causal consistency metrics

🏆 Acceptance Evidence

IEEE Acceptance

This work has been peer-reviewed and accepted for publication in IEEE Access, highlighting its contribution to reliable medical AI research.

📖 Official Publication

🏆 IEEE Access (Early Access, February 2026)

🔗 Official IEEE Xplore Link
https://ieeexplore.ieee.org/document/11408775

📌 DOI:
10.1109/ACCESS.2026.3667575

🧠 Why LymphAware Matters

Medical AI systems must be:

✔ Robust across scanners and hospitals
✔ Grounded in biological morphology
✔ Clinically interpretable
✔ Stable under domain shift

LymphAware moves the field closer to trustworthy computational pathology by addressing shortcut bias at the representation level, rather than relying solely on dataset curation or domain labels.

🚀 Training LymphAware

We provide a clean, reproducible PyTorch pipeline located in the src/ directory for training LymphAware across multi-center lymphoma datasets.

The framework is backbone-agnostic and supports:

🧠 ResNet (18 / 50 / 152)
🌿 DenseNet (121)
🔭 Vision Transformers (ViT-L/16)
🏥 Multi-center domain training (Centers A–E)
🎨 Artifact-shift augmentation for shortcut exposure
📈 AUC and FPR evaluation

📂 Project Structure

LymphAware/
│
├── src/
│   ├── train_lymphaware.py
│   ├── models/
│   ├── datasets/
│   ├── losses/
│   └── utils/
│
├── data/
│   ├── CenterA/
│   ├── CenterB/
│   ├── CenterC/
│   ├── CenterD/
│   └── CenterE/
│
└── outputs/

Each center directory should contain class folders:

CenterA/
    CLL/
    FL/
    MCL/

⚙️ Installation

git clone https://github.com/kaopanboonyuen/LymphAware.git
cd LymphAware

conda create -n lymphaware python=3.10
conda activate lymphaware

pip install -r requirements.txt

▶️ Training Example

Train on a specific center (e.g., Center A):

python src/train_lymphaware.py \
    --train_dir data/CenterA/train \
    --val_dir data/CenterA/test \
    --backbone resnet50 \
    --epochs 100 \
    --batch_size 16 \
    --lr 3e-4

🔬 Training with Vision Transformer (Best Performance)

python src/train_lymphaware.py \
    --train_dir data/CenterA/train \
    --val_dir data/CenterA/test \
    --backbone vit_large_patch16_224 \
    --epochs 100

💾 Outputs

Training artifacts will be saved to:

outputs/
    best_model.pth

The script automatically:

✅ Tracks validation AUC ✅ Computes False Positive Rate (FPR) ✅ Saves the best checkpoint ✅ Supports GPU acceleration

🧪 Multi-Center Reproduction (Centers A–E)

To reproduce the paper results:

Train a model per center
Evaluate cross-domain performance
Average metrics across runs

Example loop:

for CENTER in CenterA CenterB CenterC CenterD CenterE
do
  python src/train_lymphaware.py \
      --train_dir data/${CENTER}/train \
      --val_dir data/${CENTER}/test \
      --backbone vit_large_patch16_224
done

⭐ Research Tips (From the Paper)

For best performance reported in IEEE Access:

Backbone: ViT-L/16
Epochs: 100
Optimizer: AdamW
Learning rate: 3e-4
Image size: 224 × 224
Loss weight (orthogonality): 0.1

🧠 Why This Training Matters

Unlike standard pipelines, LymphAware training:

Disrupts shortcut bias during representation learning
Encourages morphology-grounded predictions
Improves robustness across scanners and institutions
Produces clinically meaningful attribution behavior

The model learns cancer morphology — not acquisition artifacts.

If you find this work useful, please ⭐ star the repository.

🙏 Acknowledgement

This research is supported by:

🎓 Talent Scholarship for Exceptional Ability 🏫 College of Computing, Khon Kaen University

🌟 Final Note

LymphAware learns the cancer — not the confounders.

By enforcing morphology-grounded representations and suppressing shortcut bias, we aim to build AI systems that clinicians can truly trust.

⭐ If you find this project useful, please consider starring the repository!

📚 BibTeX Citation

@article{panboonyuen2026lymphaware,
  author    = {Teerapong Panboonyuen},
  title     = {LymphAware: Domain-Aware Bias Disruption for Reliable Lymphoma Cancer AI Diagnosis},
  journal   = {IEEE Access},
  year      = {2026},
  pages     = {1--1},
  doi       = {10.1109/ACCESS.2026.3667575},
  publisher = {IEEE}
}

If you use this work in your research, please cite the official IEEE version via the DOI above.

This site is open source. Improve this page.