Biased CAMELYON16 Bias Analysis

• The paper introduces a formal framework using the generalized Rayleigh quotient to isolate intrinsic bias signals from background artifacts in CAMELYON16.
• It presents KlotskiNet and BDDA methods that employ eigenanalysis and Grad-CAM visualizations to quantify bias directions in gigapixel WSIs.
• The Ada-ABC debiasing framework demonstrates improved balanced performance by mitigating shortcut-induced bias, enhancing model robustness in tumor detection.

Biased CAMELYON16 refers to the introduction, identification, and remediation of implicit or explicit bias in the CAMELYON16 medical imaging dataset—a curated set of gigapixel whole-slide images (WSIs) of lymph node sections used for metastatic tumor detection. The concept encompasses (i) intrinsic dataset biases, such as background artifacts or scanner-dependent features, and (ii) artificially injected spurious correlations, such as associating slide acquisition center with tumor presence. These biases can severely degrade the generalization of machine learning models trained on such data and motivate systematic approaches for bias discovery and debiasing.

1. Intrinsic Bias Identification in CAMELYON16

CAMELYON16 WSIs exhibit background features (e.g., staining patterns, tissue folding, scanner artifacts) that are often correlated with clinical labels but do not represent true lesion information. A formal “intrinsic bias attribute” is defined as a unit vector $\varphi \in \mathbb{R}^K$ such that the projections $v_i = \varphi^T u_i$—where $u_i$ is a background-biased embedding of slide $i$—have statistically distinct distributions for positive (lesion-present) and negative (lesion-absent) samples, independent of true lesion features. The divergence between these distributions is maximized by selecting $\varphi$ that solves the generalized Rayleigh quotient:

$$\varphi^* = \underset{\varphi:\|\varphi\|=1}{\text{argmax}} \; \frac{\varphi^T S_{PN} \varphi}{\varphi^T S_P \varphi}$$

Here, $S_P$ and $S_{PN}$ are sample covariance matrices for positive and negative groups, respectively. This formalism isolates dataset-specific signals that can be exploited by models in downstream tasks, leading to potentially spurious predictions (Zhang et al., 2022).

2. KlotskiNet and Bias Discriminant Direction Analysis (BDDA)

KlotskiNet is an architecture designed to map image tiles to embeddings and class probabilities, emphasizing background cues over lesion features. It employs a ResNet-50 backbone truncated before final pooling, feeding outputs to fully-connected layers and a softmax head. During training, only background tiles of each slide (determined via lesion masks) are considered; the tile with maximum model output confidence is selected per slide, and cross-entropy is minimized:

$$L_i(\theta) = -\left[ \mathbb{I}_{y_i=+1} \log q^+_{r^*} + \mathbb{I}_{y_i=-1} \log q^-_{r^*} \right]$$

Aggregated over all slides, this encourages memorization of dataset-specific background artifacts. After training, embeddings $u_i$ for each slide are collected and partitioned by label. BDDA finds principal bias directions $\varphi_1, \ldots, \varphi_k$ by solving a generalized eigenproblem, enforcing conjugate orthogonality for higher-order directions. These directions illuminate uncorrelated bias features within the dataset (Zhang et al., 2022).

3. Artificially Injected Bias and the Ada-ABC Debiasing Framework

Beyond intrinsic bias, CAMELYON16 can be deliberately “biased” by constructing a training split where tumor patches are predominantly sourced from one center (e.g., Radboud), and normal patches from another (e.g., UMC Utrecht)—creating a dominant shortcut between scanner domain and label, independent of biological reality (Luo et al., 2024). The Ada-ABC (Adaptive Agreement from a Biased Council) framework addresses such cases without explicit bias attribute labels.

Ada-ABC comprises:

• A “biased council” (ensemble of K heads sharing a feature extractor) trained with Generalized Cross Entropy (GCE), which specializes in “easy” (shortcut-rich) predictions;
• A debiasing model trained with an adaptive agreement objective: it agrees with the council on likely correct samples and disagrees where the council predicts incorrectly (i.e., where shortcut-induced bias is likely).

With notation:

$$\mathcal{L}_{\rm Ada}(f_D, f_B; x, y) = p \,\mathcal{L}_{\rm agr}(f_D; x, y) + (1 - p)\,\mathcal{L}_{\rm opp}(f_D, f_B; x)$$

where $p$ is the mean tumor score from the council, $\mathcal{L}_{\rm agr}$ is standard cross-entropy, and $\mathcal{L}_{\rm opp}$ encourages output disagreement. Training fuses these losses so that the debiasing model learns genuine tumor morphology rather than shortcut features (Luo et al., 2024).

4. Application Pipeline and Quantitative Results

Preprocessing for both intrinsic and injected bias workflows includes:

• Tiling WSIs into patches (e.g., 256×256 or 224×224 at 0.5 µm/pixel)
• Color normalization with Macenko's method to reduce stain variability
• Background/foreground segmentation to isolate non-lesion regions

Bias evaluation employs metric splits:

• AUC (Area Under the Curve) on bias-aligned (shortcut-exploiting) and bias-conflict (shortcut-invalid) test samples
• Balanced AUC, accuracy, sensitivity, specificity

Experimental results with Ada-ABC on a strongly biased CAMELYON16 split (correlation $\rho = 0.95$) yield substantial improvements:

Method	AUC_aligned	AUC_conflict	Balanced AUC	Acc	Sens	Spec
ERM	0.98 ± 0.01	0.62 ± 0.03	0.80 ± 0.02	0.84	0.85	0.83
LfF	0.89 ± 0.02	0.75 ± 0.04	0.82 ± 0.03	0.84	0.81	0.87
JTT	0.92 ± 0.02	0.78 ± 0.03	0.85 ± 0.02	0.86	0.84	0.88
PBBL	0.90 ± 0.03	0.81 ± 0.02	0.86 ± 0.02	0.87	0.85	0.89
Ada-ABC	0.92 ± 0.01	0.90 ± 0.02	0.91 ± 0.01	0.90	0.89	0.91

Ada-ABC achieves nearly equivalent performance on bias-aligned and conflict splits and improves balanced AUC by 0.11 compared to standard empirical risk minimization (Luo et al., 2024).

5. Visualization and Validation of Bias Attributes

Bias directions found by BDDA are visualized:

• Histograms of $v_{i,j} = \varphi_j^T u_i$ for positives and negatives reveal the degree of separation;
• Grad-CAM overlays on the background tiles driving high $v_{i,j}$ values elucidate the actual image regions responsible for bias—frequently slide-level artifacts such as tissue edges or scanner-specific patterns.

For injected bias, Ada-ABC demonstrates restoration of true morphological feature learning by forcing disagreement with council predictions rooted in shortcuts. A plausible implication is improved robustness and fairness, with sensitivity/specificity balanced across scanner domains (Zhang et al., 2022, Luo et al., 2024).

6. Limitations, Practical Considerations, and Future Directions

Key considerations include:

• Computational cost: Tiling gigapixel WSIs at fine resolution yields millions of patches. Strategies are required to mitigate memory and I/O bottlenecks, such as tissue pre-filtering and parallel loading.
• Balanced groups: BDDA effectiveness depends on sufficient samples per class; severe class imbalance necessitates subsampling or regularized covariance estimation.
• Mask accuracy: Precise foreground removal mandates reliable tumor masks, often challenging in weakly annotated datasets.
• Number of bias directions ($k$): Selecting too many directions introduces noise and diminishes signal; dataset-specific eigenvalue spectra should guide $k$.
• Fundamental limitations: BDDA finds linear bias directions; nonlinear biases may require kernel extensions or contrastive losses. Ada-ABC depends on the biased council capturing a dominant shortcut; weakness or multiplicity of actual biases may reduce efficacy.

Both frameworks—KlotskiNet+BDDA and Ada-ABC—advance automated, quantitative bias characterization and remediation in CAMELYON16, serving to improve the reliability and fairness of medical imaging models. Future work includes extending these methods to multiclass settings and further exploring nonlinear bias discovery (Zhang et al., 2022, Luo et al., 2024).