On the Interplay of Human-AI Alignment,Fairness, and Performance Trade-offs in Medical Imaging

TL;DR

The paper addresses fairness gaps and robustness in medical imaging by systematically studying Human-AI alignment in chest X-ray disease classification. It introduces a multi-dimensional experimental design with five alignment levels, multicenter training, and diverse OOD data, evaluated using subgroup fairness (AUC gaps across sex and age) and standard performance metrics. A ViT-based architecture with a cross-modal Vision-Language Model and an Attention Aligner loss combines disease prompts with radiologist-attended regions, formalized through $\\mathcal{L}_{AL}$ and $\\mathcal{L}_{CE}$ in $\\mathcal{L}_{\\text{total}} = \\mathcal{L}_{CE} + \\mathcal{L}_{AL}$. The results show that alignment generally reduces fairness gaps and enhances OOD generalization, especially in low-data settings, but excessive or randomized alignment can incur trade-offs, underscoring the need for calibrated alignment strategies for fair, robust medical AI systems.

Abstract

Deep neural networks excel in medical imaging but remain prone to biases, leading to fairness gaps across demographic groups. We provide the first systematic exploration of Human-AI alignment and fairness in this domain. Our results show that incorporating human insights consistently reduces fairness gaps and enhances out-of-domain generalization, though excessive alignment can introduce performance trade-offs, emphasizing the need for calibrated strategies. These findings highlight Human-AI alignment as a promising approach for developing fair, robust, and generalizable medical AI systems, striking a balance between expert guidance and automated efficiency. Our code is available at https://github.com/Roypic/Aligner.

Figures (6)

• Figure 1: Experimental setup to assess the impact of Human-AI alignment on fairness. Models trained on multicenter training datasets are trained without Human-AI alignment (i.e., baseline), and with various degrees of Human-AI alignment, and their fairness gap and classification performance metrics are assessed across two demographic groups on out-of-domain datasets across three different classification tasks. Additionally, the impact of Human-AI attention on low-data regimes and when alignment is randomized are further evaluated.
• Figure 2: Human-AI Alignment flow chart adapted from luo2024dwarf. The approach is based on Visual-Language-Model (VLM) fusing image and language embeddings via cross-attention. The model is trained sequentially for each class per epoch, with the disease name as a prompt (e.g., "Edema"). Two projector heads are used to (i) optimize Human-AI alignment, and (ii) perform disease classification.
• Figure 3: Fairness-performance trade-off for age and sex groups across five levels of Human-AI alignment. Blue points represent non-aligned models ($0\%$), while red-shaded points ($25\%$–$100\%$) indicate increasing alignment. Error bars show variability, and red-shaded ellipses highlight the trends. Fairness improves up to $75\%$ alignment but degrades at $100\%$, suggesting an overconstraining effect.
• Figure 4: Fairness-performance trade-off results under randomized Human-AI alignment (green) for sex group demographics, compared to a baseline model without alignment (blue), and with Human-AI alignment (red). Error bars represent variability.
• Figure 5: Fairness comparison of Human-AI aligned (red) and non-aligned (blue) models across training data ratios and performance metrics. Lower values indicate better fairness. Alignment reduces fairness gaps, especially in low-data settings