Fair Lung Disease Diagnosis from Chest CT via Gender-Adversarial Attention Multiple Instance Learning

Aditya Parikh, Aasa Feragen

Abstract

We present a fairness-aware framework for multi-class lung disease diagnosis from chest CT volumes, developed for the Fair Disease Diagnosis Challenge at the PHAROS-AIF-MIH Workshop (CVPR 2026). The challenge requires classifying CT scans into four categories — Healthy, COVID-19, Adenocarcinoma, and Squamous Cell Carcinoma — with performance measured as the average of per-gender macro F1 scores, explicitly penalizing gender-inequitable predictions. Our approach addresses two core difficulties: the sparse pathological signal across hundreds of slices, and a severe demographic imbalance compounded across disease class and gender. We propose an attention-based Multiple Instance Learning (MIL) model on a ConvNeXt backbone that learns to identify diagnostically relevant slices without slice-level supervision, augmented with a Gradient Reversal Layer (GRL) that adversarially suppresses gender-predictive structure in the learned scan representation. Training incorporates focal loss with label smoothing, stratified cross-validation over joint (class, gender) strata, and targeted oversampling of the most underrepresented subgroup. At inference, all five-fold checkpoints are ensembled with horizontal-flip test-time augmentation via soft logit voting and out-of-the-fold threshold optimization for robustness. Our model achieves a mean validation competition score of 0.685 (std - 0.030), with the best single fold reaching 0.759.

Introduction: The Challenges of Sparse Signals and Demographic Bias

Automated diagnosis from chest CT scans faces two primary obstacles: signal sparsity and demographic bias. In a 3D CT volume containing hundreds of axial slices, pathological indicators like tumors or infections may only manifest in a small fraction of the data. Conventional pooling methods often fail; mean-pooling dilutes the signal, while max-pooling is overly sensitive to noise.

Furthermore, medical datasets frequently exhibit severe imbalances. In the PHAROS-AIF-MIH Workshop's Fair Disease Diagnosis Challenge, female patients with Squamous Cell Carcinoma (SCC) were significantly underrepresented. Without intervention, AI models may learn "shortcuts"—associating diseases with gender-specific features like body morphology rather than actual pathology—leading to inequitable diagnostic performance for minority groups.

Methodology: Attention-MIL and Gender-Adversarial Training

To address these challenges, we propose a multi-stage pipeline designed for both sensitivity and equity.

Attention-based Multiple Instance Learning (MIL)

Instead of requiring expensive slice-level labels, our model treats each CT volume as a "bag" of instances. A ConvNeXt backbone extracts features from each slice, and an attention mechanism learns to assign importance weights. This allows the model to automatically focus on diagnostically relevant slices while ignoring uninformative background regions.

Gender-Adversarial Training via Gradient Reversal

To prevent the model from exploiting gender as a latent shortcut, we integrated a secondary gender classifier head. During training, a Gradient Reversal Layer (GRL) is employed. This layer forces the backbone to learn representations that are uninformative for gender classification, effectively stripping away gender-biased features and focusing the model purely on pathology.

Training and Inference Robustness

We utilized Focal Loss with label smoothing to handle difficult examples and a WeightedRandomSampler to ensure rare female SCC cases were present in every training batch. At inference, we employed Out-of-Fold (OOF) threshold optimization to determine optimal probability cut-offs for each class, further boosting performance on rare subgroups.

Dataset and Distribution Analysis

The study utilized a dataset of 889 3D chest CT scans (734 for training, 155 for validation) across four classes: Adenocarcinoma, COVID-19, Healthy, and Squamous Cell Carcinoma.

Class distribution by gender and slice count distribution per scan, illustrating the severe intersectional imbalance and volume variability.

As shown in the distribution analysis, while the overall gender split is relatively balanced (481 male, 408 female), there is a severe intersectional imbalance in the SCC class, with 91 males versus only 18 females. Additionally, the number of slices per scan varies significantly across classes, complicating the "needle in a haystack" problem.

Results and Discussion

Our framework was evaluated using the average of per-gender macro F1 scores, ensuring that performance on minority subgroups is not masked by majority group accuracy.

Validation overall macro F1 scores across five training folds, demonstrating convergence and the robustness of the ensemble approach.

The model achieved a mean validation competition score of 0.685, with the best single fold reaching 0.759. Crucially, the integration of the GRL successfully closed the fairness gap: the mean female macro-F1 (0.691) was slightly higher than the male macro-F1 (0.679), indicating that the model no longer relies on gender bias for classification. Despite these gains, Squamous Cell Carcinoma remains the most challenging class (mean F1 0.366) due to its clinical overlap with other opacities and the extreme scarcity of female training examples.

Conclusion

By combining Attention-based MIL with adversarial training, we have developed a framework that is both diagnostically sensitive and demographically fair. Future work will explore generative augmentation to synthesize data for underrepresented subgroups and semi-supervised pre-training to further enhance backbone feature extraction.

All training and inference code is publicly available at https://github.com/ADE-17/cvpr-fair-chest-ct

Summary updated

Abstract

We present a fairness-aware framework for multi-class lung disease diagnosis from chest CT volumes, developed for the Fair Disease Diagnosis Challenge at the PHAROS-AIF-MIH Workshop (CVPR 2026). The challenge requires classifying CT scans into four categories -- Healthy, COVID-19, Adenocarcinoma, and Squamous Cell Carcinoma -- with performance measured as the average of per-gender macro F1 scores, explicitly penalizing gender-inequitable predictions. Our approach addresses two core difficulties: the sparse pathological signal across hundreds of slices, and a severe demographic imbalance compounded across disease class and gender. We propose an attention-based Multiple Instance Learning (MIL) model on a ConvNeXt backbone that learns to identify diagnostically relevant slices without slice-level supervision, augmented with a Gradient Reversal Layer (GRL) that adversarially suppresses gender-predictive structure in the learned scan representation. Training incorporates focal loss with label smoothing, stratified cross-validation over joint (class, gender) strata, and targeted oversampling of the most underrepresented subgroup. At inference, all five-fold checkpoints are ensembled with horizontal-flip test-time augmentation via soft logit voting and out-of-the-fold threshold optimization for robustness. Our model achieves a mean validation competition score of 0.685 (std - 0.030), with the best single fold reaching 0.759. All training and inference code is publicly available at https://github.com/ADE-17/cvpr-fair-chest-ct
BibTeX
@misc{parikh2026fair,
  author = {Aditya Parikh and Aasa Feragen},
  title = {{Fair Lung Disease Diagnosis from Chest CT via Gender-Adversarial Attention Multiple Instance Learning}},
  year = {2026},
  eprint = {2603.12988},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.12988},
}
Figure