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

Conference: CVPR 2026 arXiv: 2603.12988 Code: GitHub Area: Medical Imaging Keywords: Fairness, Lung Disease Diagnosis, CT Classification, Multiple Instance Learning, Adversarial Training

TL;DR

A fairness-aware framework based on attention MIL and gradient reversal layers (GRL) is proposed for multi-class lung disease diagnosis from chest CT volumes, eliminating gender bias while preserving diagnostic accuracy.

Background & Motivation

Automated deep learning analysis of chest CT carries significant clinical value for lung cancer screening and COVID-19 detection; however, models may encode and amplify demographic disparities present in training data, leading to systematic unfairness toward minority groups. The Fair Disease Diagnosis Challenge held at the CVPR 2026 PHAROS-AIF-MIH Workshop requires classifying CT scans into four categories (Healthy, COVID-19, Adenocarcinoma, Squamous Cell Carcinoma), evaluated by the gender-stratified macro F1 mean:

\[P = \frac{1}{2}(\text{MacroF1}_{\text{male}} + \text{MacroF1}_{\text{female}})\]

This metric explicitly penalizes gender-unequal predictions. The paper addresses three core challenges: (1) sparse pathological signals in CT volumes (only a minority of 200+ slices contain lesions); (2) severe demographic imbalance (only 18 female vs. 91 male squamous cell carcinoma cases); and (3) gender potentially serving as a latent shortcut feature exploited by the model.

Method

Overall Architecture

An attention MIL model built on a ConvNeXt backbone, augmented with a GRL adversarial branch to enforce gender fairness. The pipeline comprises four stages: slice feature extraction, attention pooling, disease classification, and adversarial gender classification.

Key Designs

1. Attention MIL Pooling: Each CT volume is treated as a bag of slices. A ConvNeXt-Base encoder extracts a \(D\)-dimensional embedding \(h_i = f_{\text{enc}}(x_i)\) per slice; a two-layer MLP attention network then assigns importance weights \(w_i = \frac{\exp(s_i)}{\sum_j \exp(s_j)}\) to each slice, and a weighted aggregation yields the scan-level representation \(H = \sum_i w_i h_i\). Zero-padded positions are masked via a binary mask. The core motivation is to enable the model to automatically learn the importance of diagnostically relevant slices without slice-level annotations.

2. Gradient Reversal Layer (GRL) Adversarial Debiasing: A gender classifier is appended to the scan embedding \(H\), with the GRL reversing and scaling gradients during backpropagation: \(z_{\text{gen}} = c(\mathcal{R}_\lambda(H))\). The training objective is \(\mathcal{L} = \mathcal{L}_{\text{disease}} + \lambda_{\text{adv}} \mathcal{L}_{\text{gender}}\). The design motivation is that even without explicit gender inputs, the backbone may encode gender information from CT acquisition parameters and body morphology as spurious correlations.

3. Subgroup Oversampling and Stratified CV: A WeightedRandomSampler substantially increases the sampling weight of female squamous cell carcinoma cases (Female G), ensuring the rarest subgroup is represented in every batch. Five-fold cross-validation is stratified on the joint (class, gender) key, guaranteeing representation of all eight subgroups in each fold.

Loss & Training

• Focal Loss + Label Smoothing: \(\mathcal{L}_{\text{disease}} = -\alpha(1-p_t)^\gamma \log \tilde{p}_t\), with \(\gamma=2\), \(\alpha=0.25\), smoothing \(\varepsilon=0.1\), concentrating gradients on hard samples and scarce subgroups.
• Two-Stage Fine-Tuning: Epochs 1–5 freeze the backbone (LR \(10^{-3}\)); from Epoch 6 onward the backbone is unfrozen (backbone LR \(10^{-5}\), heads LR \(10^{-4}\)) with cosine annealing.
• Gradient Accumulation: \(K=4\) steps, yielding an effective batch size of 16 volumes.
• At most \(M=32\) slices per volume; random sampling during training and uniform sampling during inference.

Key Experimental Results

Dataset

889 3D chest CT scans (734 train / 155 validation), four classes: Adenocarcinoma (300), COVID-19 (240), Healthy (240), Squamous Cell Carcinoma (109). The overall gender distribution is relatively balanced (481 male / 408 female), but only 18 female vs. 91 male SCC cases; volume depth varies widely (20–800+ slices per scan).

Main Results

Dataset	Metric	Ours	Best Single Fold	Notes
889 CT scans (734 train)	Competition Score P	0.685 ± 0.030	0.759 (Fold 1)	5-fold mean
Same	Male macro-F1	0.679 ± 0.068	0.754	Gender gap reduced after GRL
Same	Female macro-F1	0.691 ± 0.030	0.722	Female F1 slightly higher than male

Ablation Study

Configuration	Key Metric	Notes
Mean Pooling (Baseline)	Low	Tumor signal diluted by healthy slices
+ Max Pooling	Improved	Restores detection of sparse tumor slices
+ Attention-MIL	Further improved	Learns to ignore blank lung regions
+ Subgroup Oversampling	Significant gain in Female F1	Prevents minority class collapse
+ GRL	P = 0.685	Closes fairness gap; male and female performance equalized

Key Findings

• GRL successfully narrows the gender fairness gap: female macro-F1 (0.691) slightly exceeds male (0.679).
• Squamous cell carcinoma (SCC) remains the most challenging class (F1 = 0.366 ± 0.083), limited by data scarcity and clinical overlap.
• OOF threshold optimization maintains strict gender fairness (Male F1 0.679 vs. Female F1 0.688).

Highlights & Insights

• The fairness problem is decomposed into three distinct failure modes (sparse signals, demographic imbalance, latent shortcut features), each addressed by a dedicated module.
• GRL adversarial training offers an elegant solution for eliminating gender bias in feature space, more thorough than simple data balancing.
• OOF threshold optimization avoids overfitting to the validation set and serves as a practical post-hoc fairness calibration technique.

Limitations & Future Work

• Female SCC data remains extremely scarce (only 18 cases); oversampling cannot fully compensate.
• Generative data augmentation (e.g., diffusion-based CT synthesis for scarce subgroups) is unexplored.
• The fairness constraint could be extended to a stronger fairness-constrained optimization formulation.
• Attention visualization and clinical interpretability are not discussed in depth.
• The 5-fold ensemble at inference incurs high computational cost, requiring each test sample to pass through all five models.

Related Work & Insights

• The GRL domain adaptation method of Ganin & Lempitsky (2015) is elegantly transferred to the fairness setting.
• The Attention-MIL framework of Ilse et al. (2018) is well-suited to handling sparse signals in CT volumes.
• The combination of Focal Loss (Lin et al., 2017) and Label Smoothing is friendly to scarce subgroups.
• The inference-time combination of 5-fold ensemble, TTA, and OOF threshold optimization yields robust challenge performance.

Rating

• Novelty: ⭐⭐⭐ All components are established methods; however, their integration for fair diagnostic classification is valuable.
• Experimental Thoroughness: ⭐⭐⭐ 5-fold CV and ablation are reasonably complete, though the dataset is small and ablations are primarily qualitative.
• Writing Quality: ⭐⭐⭐⭐ Problem decomposition is clear, motivation is well-articulated, and the challenge paper format is well-structured.
• Value: ⭐⭐⭐ A practical framework for fairness in medical AI, though it primarily constitutes a challenge solution.

Additional Notes

This paper presents a competition entry for the CVPR 2026 PHAROS-AIF-MIH Workshop Challenge, achieving a robust competition score of 0.685. The overall methodology is engineering-oriented; nonetheless, the multi-level fairness constraint design — oversampling at the data level, GRL at the feature level, and OOF thresholding at the decision level — provides a reusable toolkit for fairness in medical AI.

Related Papers

• [CVPR 2026] Robust Fair Disease Diagnosis in CT Images
• [CVPR 2026] MIL-PF: Multiple Instance Learning on Precomputed Features for Mammography Classification
• [CVPR 2026] Every Error has Its Magnitude: Asymmetric Mistake Severity Training for Multiclass Multiple Instance Learning
• [CVPR 2026] EMAD: Evidence-Centric Grounded Multimodal Diagnosis for Alzheimer's Disease
• [CVPR 2026] Cell-Type Prototype-Informed Neural Network for Gene Expression Estimation from Pathology Images