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

Aditya Parikh; Aasa Feragen

cs.CVMar 13, 2026

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

Authors: Aditya Parikh, Aasa Feragen

Organizations: Technical University of Denmark (DTU), Kongens Lyngby, Denmark

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

Figures & tables

Figure 1: Representative axial CT slices for each diagnostic category. (a) Healthy lungs exhibit clear parenchyma. (b) COVID-19 infection presents with characteristic bilateral ground-glass opacities distributed across the lung fields. (c) Adenocarcinoma and (d) Squamous Cell Carcinoma represent the two malignant categories, each presenting as distinct focal lung lesions. The subtle visual differences between disease classes, and the fact that abnormalities occupy only a small fraction of slices within a full CT volume, motivate our attention-based MIL approach.

Panel of Figure 2 — Figure 2: Dataset characteristics. (top) Distribution of scans by class and gender, highlighting the severe intersectional scarcity of female Squamous Cell Carcinoma (SCC) cases. (bottom) Variance in volumetric depth across classes. The extreme fluctuation in slices per scan (ranging from under 20 to over 800) necessitates our flexible Attention-MIL formulation.

Fold	$P$	M-F1	F-F1	F1-A	F1-G
0	0.698	0.673	0.722	0.807	0.258
1	0.727	0.754	0.699	0.796	0.378
2	0.674	0.658	0.690	0.692	0.500
3	0.688	0.743	0.634	0.803	0.303
4	0.637	0.565	0.709	0.681	0.389
Mean	0.685	0.679	0.691	0.756	0.366

Table 1: Per-fold validation metrics. The competition score

P

is the average of male and female macro-F1. Best fold highlighted in bold .

Figure 3: Weights & Biases validation curves across all 5 folds. The integration of the Gradient Reversal Layer stabilized the initially highly variant training dynamics.

Design Choice	Primary Challenge Addressed	Observed Improvement
Mean Pooling (Baseline)	N/A	Model struggled to detect small tumors; signals diluted by healthy slices.
+ Max Pooling	Localized tumor signal dilution in 3D volumes.	Recovered ability to trigger positive predictions from sparse tumor slices.
+ Attention-MIL	Noise from uninformative background and boundary slices.	Improved spatial robustness by learning to actively ignore empty lung regions.
+ Subgroup Oversampling	Extreme intersectional scarcity (e.g., only 5 Female SCC scans).	Prevented minority class collapse and substantially lifted Female Macro-F1.
+ Gradient Reversal Layer	Entanglement of clinical tumor features with gender features.	Closed the fairness gap, equating Male and Female performance ( $P=0.685$ ).

Table 2: Qualitative ablation study detailing the architectural and methodological evolution of the pipeline. Each component was introduced to address a specific observed bottleneck.

Model	$P$	M-F1	F-F1	F1-A	F1-G	F1-Cov
OOF Global Mean	0.683	0.679	0.688	0.755	0.366	0.813
OOF Global $\pm$	0.032	0.066	0.029	0.056	0.083	0.070
Ens. All+TTA (Test)	–	–	–	–	–	–

Table 3: Out-of-Fold (OOF) evaluation metrics representing the true global performance of the 5-fold ensemble with threshold optimization. Ensemble test metrics to be updated upon label release.

Explore similar work

Robust Fair Disease Diagnosis in CT Images
Apr 8, 2026Justin Li, Daniel Ding, Asmita Yuki Pritha +3Algorithmic Fairness Medical Image Classification
Carmel High School, Carmel, Indiana, USA · Capstone School Dhaka, Bangladesh · Clarkstown High School South, West Nyack, New York, USA +2
Clinical Priors Guided Lung Disease Detection in 3D CT Scans
Mar 16, 2026Kejin Lu, Jianfa Bai, Qingqiu Li +5Medical Image Classification 3D Medical Imaging
College of Computer Science and Artificial Intelligence, Shanghai Key Laboratory of Intelligent Information Processing, Fudan University · University of Oxford · The Hong Kong University of Science and Technology
Towards Fair and Robust Volumetric CT Classification via KL-Regularised Group Distributionally Robust Optimisation
Mar 16, 2026Samuel Johnny, Blessed Guda, Goodness Obasi +2Medical Image Classification Distributionally Robust Optimization
Carnegie Mellon University Africa, Kigali, Rwanda · Carnegie Mellon University, Pittsburgh, USA

Apr 8, 2026J/K move · Enter open · S save

Robust Fair Disease Diagnosis in CT Images

Justin Li, Daniel Ding, Asmita Yuki Pritha, Aryana Hou+2

Carmel High School, Carmel, Indiana, USA · Capstone School Dhaka, Bangladesh · Clarkstown High School South, West Nyack, New York, USA+2

Algorithmic Fairness Medical Image Classification Class-Imbalanced Learning

Open paper AskPDF arXiv

Automated diagnosis from chest CT has improved considerably with deep learning, but models trained on skewed datasets tend to perform unevenly across patient demographics. However, the situation is worse than simple demographic bias. In clinical data, class imbalance and group underrepresentation often coincide, creating compound failure modes that neither standard rebalancing nor fairness corrections can fix alone. We introduce a two-level objective that targets both axes of this problem. Logit-adjusted cross-entropy loss operates at the sample level, shifting decision margins by class frequency with provable consistency guarantees. Conditional Value at Risk aggregation operates at the group level, directing optimization pressure toward whichever demographic group currently has the higher loss. We evaluate on the Fair Disease Diagnosis benchmark using a 3D ResNet-18 pretrained on Kinetics-400, classifying CT volumes into Adenocarcinoma, Squamous Cell Carcinoma, COVID-19, and Normal groups with patient sex annotations. The training set illustrates the compound problem concretely: squamous cell carcinoma has 84 samples total, 5 of them female. The combined loss reaches a gender-averaged macro F1 of 0.8403 with a fairness gap of 0.0239, a 13.3% improvement in score and 78% reduction in demographic disparity over the baseline. Ablations show that each component alone falls short. The code is publicly available at https://github.com/Purdue-M2/Fair-Disease-Diagnosis.

Abstract

Explore similar work

Robust Fair Disease Diagnosis in CT Images

Clinical Priors Guided Lung Disease Detection in 3D CT Scans

Towards Fair and Robust Volumetric CT Classification via KL-Regularised Group Distributionally Robust Optimisation