Technical Summary: MedXplore

Problem Statement

Deep learning models in medical imaging currently rely heavily on large-scale annotated datasets and a "closed-world" assumption, where unlabeled data is assumed to belong to the same categories as the training set. This assumption fails in clinical practice, where new or rare diseases constantly emerge, and exhaustive disease coverage is infeasible. While Generalized Category Discovery (GCD)—the task of recognizing known classes while discovering unknown ones—has advanced in natural image domains, it remains underexplored in medical imaging.

Existing GCD methods struggle in medical contexts due to two primary factors:

1. Domain Specificity and Variability: Medical images exhibit strong cross-device variability (modalities, scanning parameters) and subtle lesion cues. Backbones pretrained on natural images often fail to capture modality-specific structural priors, leading to weak lesion representations.
2. Semantic Ambiguity and Bias: Medical images often contain high heterogeneity where disease subtypes differ only in subtle local patterns. Limited labeled data exacerbates "old-class bias," causing models to incorrectly classify novel, ambiguous samples as known (old) classes. This results in blurred decision boundaries and unstable clustering, particularly in datasets like Kvasir where old and novel classes share nearly identical global appearances.

Methodology: MedXplore

The authors propose MedXplore, a unified framework designed to optimize medical GCD from both perceptual and decision levels. The framework builds upon the parametric GCD formulation of SimGCD but introduces two novel modules: Frequency-SNR Adaptive Attention and Consistency (FAAC) and Adaptive Cosine-Angular Margin (ACAM).

1. Perceptual Level: FAAC

FAAC addresses the issue of unreliable lesion perception and attention drift by shifting focus from spatial to frequency domains.

• Dynamic Frequency Filtering: Instead of using fixed spectral bands, FAAC employs learnable depth-wise convolutions on the real and imaginary components of the 2D Fast Fourier Transform (FFT) of feature maps. This allows the model to learn a full-spectrum filter that suppresses repetitive anatomical backgrounds while preserving task-relevant spectral components (both low-frequency structural cues and high-frequency boundary evidence).
• Global-Local Energy Contrast (GLEC): Following frequency purification, the method computes a spatial energy map. It calculates local energy fluctuations within neighborhoods and normalizes them against global spatial statistics. This GLEC activation explicitly suppresses ubiquitous background noise (e.g., UI borders, reflections) and highlights regions that structurally deviate from their surroundings, serving as reliable semantic anchors.
• Patch Consistency: FAAC selects Top-K patches based on these semantic anchors. It enforces consistency between augmented views by retaining only mutual nearest-neighbor matches with high cosine similarity. This generates a patch consistency loss ($L_{pc}$) that guides the backbone to learn texture-agnostic, lesion-sensitive representations.
• Note: FAAC is an auxiliary training module and is discarded during inference, preserving the model's architectural simplicity.

2. Decision Level: ACAM

ACAM addresses the ambiguity of category boundaries by adaptively adjusting classification margins based on sample reliability.

• Confidence Fusion: ACAM fuses three signals to estimate sample reliability ($u_b$): attention confidence (semantic saliency from FAAC), weighted-matching confidence (cross-patch consistency), and gap-based confidence (separability between positive and negative classes).
• Adaptive Margins: Based on the reliability score $u_b$, ACAM dynamically generates an angular margin ($m_{a,b}$) for positive classes and a cosine margin ($m_{c,b}$) for suppressing hard negatives.
  • High-confidence samples receive stronger geometric constraints to sharpen class boundaries.
  • Low-confidence or ambiguous samples receive gentler updates to avoid over-penalization.
• This mechanism balances intra-class compactness and inter-class separability, specifically reducing the tendency to absorb ambiguous novel samples into dominant old classes.

The total objective function combines representation learning, parametric classification, patch consistency, and the adaptive margin loss:
$$\mathcal{L} = \mathcal{L}_{rep} + \mathcal{L}_{cls} + \alpha \cdot \mathcal{L}_{pc} + \beta \cdot \mathcal{L}_{ACAM}$$

Key Contributions

1. MedXplore Framework: A unified approach for unbiased and reliable generalized category discovery in medical imaging that jointly models perceptual and discriminative cues.
2. FAAC Module: Introduces dynamic frequency purification and statistics-based semantic anchor extraction. It overcomes the limitations of fixed spectral filters and spatial attention by learning full-spectrum responses to highlight subtle lesion deviations.
3. ACAM Module: Proposes an adaptive margin mechanism that balances intra-class compactness and inter-class separability based on sample difficulty and confidence, mitigating old-class bias.
4. State-of-the-Art Performance: Demonstrates significant improvements across multiple medical benchmarks, with a specific focus on reducing false-old errors in highly ambiguous scenarios.

Experimental Results

The authors evaluated MedXplore on four datasets: NCT-CRC-HE-100K, OrganAMNIST, OrganCMNIST, and Kvasir.

• Overall Performance: MedXplore achieved a new state-of-the-art across all datasets. It showed an average 8.5% gain in "All" accuracy over the strongest competing methods.
• Kvasir Benchmark: The most significant improvement was observed on the Kvasir dataset, where MedXplore achieved 91.7% accuracy (All), surpassing the previous best (LegoGCD at 75.2%) by 16.5%.
• Bias Reduction: On Kvasir, MedXplore reduced false-old errors (misclassifying novel classes as old) from 14.50% (SimGCD) to 0.80%. This demonstrates the framework's robustness under severe old-new ambiguity.
• Ablation Studies:
  • Adding only FAAC improved accuracy significantly (e.g., +12.1% on OrganAMNIST), proving the value of frequency-based semantic anchors.
  • Adding only ACAM also improved performance, confirming the efficacy of adaptive margins.
  • The combination of both modules yielded the best results, confirming their complementary nature.

Significance and Claims

The paper claims that MedXplore addresses the critical gap in medical GCD by moving beyond the limitations of natural-image-based methods. By leveraging frequency-domain cues to capture subtle lesion deviations and using adaptive margins to handle semantic ambiguity, the framework enables the reliable discovery of unknown disease categories without relying on large-scale annotations.

The authors emphasize that the design mitigates feature bias through learnable full-spectrum filtering rather than static assumptions, stabilizing representations in heterogeneous medical data. While the current benchmarks focus on lesions with localized discriminative cues, the authors note that evaluating robustness on diffuse, low-frequency pathologies remains a direction for future work. The method is presented as a step toward open-world recognition in clinical settings where new conditions continuously emerge.