Clinically Aware Synthetic Image Generation for Concept Coverage in Chest X-ray Models

Technical Summary: CARPA - Clinically Aware Radiographic Perturbation Augmentation

Problem Statement
Deep learning models for chest X-ray (CXR) diagnosis, while achieving strong performance on standard benchmarks, face significant limitations in reliability and generalization due to incomplete coverage of clinically meaningful concept combinations in public training datasets (e.g., MIMIC-CXR). While standard data augmentation (rotation, flipping) increases sample diversity within the original data distribution, it does not introduce new clinical semantics. Synthetic image generation via GANs or diffusion models has been explored to expand training distributions; however, existing methods often operate at the image level without explicit reasoning about clinical semantics. Consequently, synthetic images may increase visual diversity without improving coverage of specific, clinically relevant concept combinations (e.g., rare or co-occurring patterns), or they may remain clinically ambiguous.

Previous work, CoRPA, attempted to address this by perturbing clinical concept vectors to generate synthetic CXRs. However, CoRPA lacks anatomical constraints, leading to synthetic images with structural inconsistencies (e.g., altered lung shapes or skeletal structures) that differ substantially from their source images. This lack of anatomical grounding limits the clinical utility of the generated data for improving model reliability.

Methodology: CARPA Framework
The authors propose CARPA (Clinically Aware Radiographic Perturbation Augmentation), a framework designed to expand clinically meaningful concept coverage while preserving anatomical correspondence to real radiographs. CARPA operates through three distinct stages:

1. Concept Encoding: Each image-report pair is encoded into a binary clinical concept vector based on an expert-defined vocabulary (e.g., "lung mass," "adenopathy"), distinguishing these specific findings from high-level diagnostic labels (e.g., "Suspicious Malignancy").
2. Targeted Concept Perturbation: The concept vector is modified using three clinically motivated strategies:
   • Intra-Class: Modifies concepts within a diagnostic label to reflect alternative valid manifestations of the same condition.
   • Insertion: Adds concepts from an additional pathology label while retaining original labels to simulate multi-pathology presentations.
   • Deletion: Removes concepts associated with a selected pathology, simulating a clinically plausible missed diagnosis or an unremarkable scan.
3. Anatomy-Preserving Image Editing: Unlike full-image synthesis, CARPA converts the perturbed concept vector into a concise clinical prompt and uses RadEdit (a latent diffusion model combined with BioViL-T and SDXL-VAE) to apply spatially localized edits to the original source image. This process expresses the intended diagnostic changes while strictly preserving global anatomical structure (e.g., lung fields, bony anatomy).

Experimental Setup
The authors evaluated CARPA across seven convolutional backbone architectures (ConvNeXt-Base/Small, DenseNet-161, EfficientNet-V2-S, ResNet-50/101, ResNeXt-101) using the MIMIC-CXR dataset (restricted to PA views).

• Training: Models were fine-tuned on stratified subsets of 10,000 synthetic images generated by either CoRPA or CARPA.
• Evaluation: Performance was tested on a held-out MIMIC-CXR benchmark. Metrics included macro-averaged AUROC, AUPRC, and F1.
• Reliability Analysis: Predictive uncertainty (entropy) and calibration (Expected Calibration Error - ECE) were measured.
• Image Quality: Structural fidelity was assessed via Structural Similarity Index (SSIM). Semantic uncertainty was evaluated using LLaVA-Rad to generate free-text descriptions from synthetic images, which were then mapped to concept vectors and compared against ground-truth perturbations (measuring Jaccard similarity, Hamming distance, and over/under-prediction rates).
• Expert Validation: Two independent radiologists evaluated subsets of synthetic images for realism and clinical agreement with proposed findings.

Key Results

• Classification Performance: Fine-tuning on CARPA-generated images consistently improved macro-AUPRC and macro-F1 across all architectures, indicating better precision-recall trade-offs under class imbalance. In contrast, fine-tuning on CoRPA images consistently degraded performance across all metrics.
• Uncertainty and Calibration: CARPA fine-tuning reduced predictive entropy (increasing confidence) and improved calibration (lower ECE) for most architectures. CoRPA fine-tuning increased both entropy and ECE, reflecting higher uncertainty and poorer calibration.
• Structural Fidelity: CARPA achieved a mean SSIM of 0.750 ± 0.031, significantly higher than CoRPA's 0.273 ± 0.089, confirming that CARPA preserves anatomical structure much more faithfully.
• Semantic Alignment: CARPA demonstrated markedly improved semantic consistency. The mean Jaccard similarity between intended and inferred concepts was 0.609 for CARPA versus 0.168 for CoRPA. CARPA also showed substantially lower rates of over-prediction (hallucinations) and under-prediction (missed findings).
• Expert Evaluation: Radiologists judged CARPA images as real at significantly higher rates (94% and 88% for two radiologists) compared to CoRPA (86% and 62%). Disagreements with CARPA images were primarily attributed to realistic diagnostic ambiguity (e.g., "subtle" or "early" findings) rather than structural implausibility or semantic errors, which were common failure modes for CoRPA.

Significance and Claims
The paper claims that anatomically grounded concept perturbations provide a more effective and clinically meaningful approach to synthetic data augmentation for CXR models than purely image-driven or unconstrained concept-driven methods. By decoupling semantic expansion from uncontrolled structural variation, CARPA enables the systematic exploration of underrepresented clinical scenarios while maintaining radiographic plausibility.

The authors conclude that this approach not only improves downstream classification performance but also enhances model reliability by reducing predictive uncertainty and improving calibration. Crucially, the method shifts the nature of model uncertainty from gross semantic mismatches and structural errors toward clinically realistic ambiguity, thereby supporting safer clinical deployment. The work suggests that enforcing anatomical constraints is essential for synthetic data to be useful for improving the robustness of medical imaging models.