MAGE: Color-Invariant and Spatial Knowledge Distillation for Gastric Neoplasm Classification

Technical Summary: MAGE – Color-Invariant and Spatial Knowledge Distillation for Gastric Neoplasm Classification

Problem Statement
Accurate differentiation between gastric adenoma and carcinoma during endoscopy is critical for clinical decision-making but remains highly challenging. The task is hindered by high inter-class similarity, ambiguous boundaries, and the tendency of deep neural networks to rely on spurious correlations (e.g., bleeding, lighting artifacts, or background context) rather than intrinsic morphological features. Existing approaches face specific limitations:

• ROI-based methods: Often suffer from error propagation due to imperfect detection or segmentation and lose surrounding global context.
• Full-image classification: Lacks necessary spatial focus and often predicts based on background regions rather than the lesion itself.
• Color dependency: Models frequently gravitate toward domain-specific texture biases, where color acts as a fragile confounder rather than a primary diagnostic indicator.

Methodology: MAGE Framework
The authors propose MAGE (Masked Achromatic Guidance Expert), a dual-branch training framework designed to enforce color-invariant morphological learning while retaining deployability on full images without annotated masks at inference time.

1. Masked Achromatic Expert (Auxiliary Branch):

   • During training, an auxiliary "expert" branch processes a masked achromatic view of the input.
   • The input is constructed by converting the White Light Image (WLI) to grayscale and applying a ground-truth segmentation mask ($\tilde{x} = m \odot \text{Gray}(x)$), setting pixels outside the Region of Interest (ROI) to zero.
   • This branch is trained solely on structural, color-invariant features, forcing it to learn discriminative patterns based on morphology rather than color cues or background artifacts.
   • This branch utilizes privileged information (masks) and is frozen during the distillation process.

2. Main Model (Deployable Branch):

   • The main branch receives the full, raw WLI (RGB) as input and is the only component used during inference.
   • It learns to mimic the expert through a dual-objective distillation strategy:
     • Classification Logit Distillation: The main model's output logits are regularized via KL divergence against the expert's logits, transferring the expert's classification confidence.
     • Spatial Attention Transfer: The main model is regularized to align its feature map attention with the expert's. By collapsing the channel dimension and computing the squared Frobenius distance between the main model's attention map and the expert's (which is constrained to the ROI), the main model is implicitly guided to focus on the neoplasm structure within the full image.

3. Architecture Choice:

   • The framework utilizes a Convolutional Neural Network (CNN) backbone rather than Vision Transformers (ViTs). The authors argue that CNNs provide necessary inductive biases (translation invariance, locality) suitable for the dataset size and ensure precise spatial alignment required for the attention loss calculation.

4. Loss Function:
   The total loss $L$ combines:

   • Weighted Binary Cross-Entropy (BCE) against ground truth labels.
   • Temperature-scaled KL divergence between expert and main model logits.
   • Attention regularization term (Frobenius distance between attention maps).

Key Contributions

1. Training-Only Auxiliary Expert: Introduction of a masked achromatic expert that promotes color-invariant morphological learning by suppressing background context and color information during training.
2. Spatial Distillation Framework: A mechanism to transfer ROI-centric structural cues and spatial attention from the expert to a deployable full-image classifier, enabling real-time inference without requiring segmentation masks.
3. Performance and Robustness: Demonstration of consistent improvements over strong detection-based (e.g., YOLO) and classification-based (e.g., Swin-Transformer, GastroNet) baselines, alongside enhanced explainability and robustness to color perturbations.

Experimental Results
Experiments were conducted on a private gastric endoscopy dataset (Seoul St. Mary's Hospital) and an external colon dataset (PICCOLO).

• Classification Performance: MAGE achieved an AUC of 0.786 and F1 of 0.733 on the gastric test set, outperforming baselines including GastroNet-5M (AUC 0.744), Swin-Tiny (AUC 0.719), and YOLO-based pipelines.
• Explainability: MAGE demonstrated superior alignment with ground-truth masks in Grad-CAM visualizations.
  • Attention-in-Mask (AiM): MAGE achieved 0.5303 (vs. 0.3176 for baseline EffNet).
  • Pointing Game Accuracy (PGA): MAGE achieved 0.7324 (vs. 0.4355 for baseline).
  • This indicates the model relies more on neoplasm morphology than spurious global patterns.
• Color Robustness: Under synthetic hue shifts, MAGE showed significantly less performance degradation compared to EfficientNet and GastroNet-pretrained ResNet50 baselines, confirming the efficacy of color-invariant training.
• Cross-Domain Generalization: On the external PICCOLO dataset, MAGE achieved the highest AUC (0.9434), suggesting the learned morphological features are transferable across different anatomical domains.
• Ablation Studies: Results confirmed that the asymmetric design (Grayscale Expert + RGB Main) is critical. The expert performed better on grayscale inputs, while the main model required full RGB input for optimal deployment.

Significance and Claims
The paper claims that MAGE addresses the "shortcut learning" problem in gastric neoplasm classification by explicitly forcing the model to learn structural, color-invariant features through a privileged training signal. By decoupling the training requirement for masks from the inference requirement, MAGE offers a practical solution that improves both diagnostic accuracy and clinical interpretability. The authors emphasize that their method provides not only superior classification metrics but also interpretable attention maps, which are essential for clinical reliability.

Limitations
The authors note that the study focuses on classification and robustness without training a dedicated detection/segmentation module, which may limit direct localization outputs in certain clinical workflows. Future work is suggested to leverage the learned classifier for lightweight localization (e.g., CAM-based) and to conduct broader multi-center validation.