PrefPaint: Enhancing Medical Image Inpainting through Expert Human Feedback

1. Problem Statement

The paper addresses the critical challenge of generating anatomically accurate synthetic medical images, specifically for colorectal polyps, to train clinical AI models.

• The Gap: While deep learning models (like GANs and Diffusion models) excel at general image inpainting, they often fail in medical contexts. Standard models trained on unannotated or synthetic data frequently produce anatomically incorrect structures (e.g., unrealistic tumor shapes or textures).
• The Consequence: Inaccurate synthetic data leads to "false patterns" in downstream diagnostic AI, potentially causing missed diagnoses (polyps are missed in 14–30% of colonoscopies).
• The Bottleneck: Traditional methods to align models with human preferences, such as Reinforcement Learning from Human Feedback (RLHF), require training intermediate reward models. This process is computationally expensive, requires massive datasets, and is often infeasible for resource-constrained clinical institutions.

2. Methodology: PrefPaint System

The authors propose PrefPaint, an interactive system that integrates expert human feedback directly into the fine-tuning of Stable Diffusion Inpainting models without relying on expensive reward models.

A. Core Algorithm: D3PO (Direct Preference Optimization)

Instead of full RLHF, the system employs D3PO (a variant of Direct Preference Optimization adapted for diffusion processes).

• Mechanism: The denoising process is modeled as a multi-step Markov Decision Process (MDP).
• Training Loop:
  1. The model generates two candidate inpainted images based on a prompt and mask.
  2. A human expert (medical professional) provides binary feedback (e.g., "Good" vs. "Bad" or "Like" vs. "Dislike") on which image is more anatomically accurate.
  3. The model parameters are updated directly using a DDPO loss function based on these binary preferences.
  4. Advantage: This bypasses the need to train a separate, computationally heavy reward model, making it feasible for clinical settings with limited GPU resources.

B. Interactive Web-Based Platform

To facilitate "human-in-the-loop" training, the authors developed a responsive web interface with the following technical architecture:

• Backend: Built with Golang using the Echo framework and PostgreSQL for data management. It utilizes Server-Side Rendering (SSR) with Templ (a typed templating engine) for UI logic and Tailwind CSS for styling.
• Task Management: Uses a FIFO (First-In, First-Out) task queue with multi-threading. Heavy GPU tasks (sampling, fine-tuning) are offloaded to worker threads, preventing the UI from freezing.
• Key Feature: Model Tree Versioning: A novel HCI concept visualizing the evolutionary progression of the model.
  • The Root Node is the baseline model.
  • Child Nodes represent iterations fine-tuned by specific batches of expert feedback.
  • This allows clinicians to visualize the "provenance" of the model and track performance improvements over time.

3. Key Contributions

1. Novel Training Paradigm: The first application of D3PO to medical image inpainting, enabling direct parameter updates via binary expert feedback without reward models.
2. Resource Efficiency: The system is designed specifically for resource-constrained clinical environments, eliminating the need for massive compute clusters required by traditional RLHF.
3. Human-Centric Interface: Development of a user-friendly web platform featuring the Model Tree visualization, making it accessible to medical professionals with limited IT backgrounds.
4. High-Quality Synthetic Data: A pipeline that generates highly realistic, anatomically correct polyp images suitable for training robust clinical AI.

4. Experimental Results

The system was evaluated using the Kvasir-SEG dataset (3,842 sessile and 1,034 pedunculated polyp images).

Quantitative Metrics

The proposed method ("Ours") outperformed baselines (Stable Diffusion Inpainting, DreamBooth, SD2i) across all metrics:

• Sessile Polyps: Achieved the lowest L1/L2 errors (0.0169/0.0033) and highest SSIM (0.8883).
• Pedunculated Polyps: Similarly achieved the best scores across L1, L2, SSIM, PSNR, and LPIPS (Learned Perceptual Image Patch Similarity).
• Significance: Lower LPIPS and higher SSIM indicate superior structural fidelity and perceptual realism compared to standard diffusion models.

Qualitative & User Studies

• Medical Expert Validation: 3 medical doctors and 25+ participants (including CS students) evaluated the outputs.
• Metrics: Alignment, Realism, Smoothness, and Quality.
• Findings:
  • Medical Experts: Rated PrefPaint significantly higher (Mean Opinion Score 3.5–3.7) compared to baselines (1.6–2.0). Experts noted that PrefPaint avoided the "anatomical artifacts" common in other methods.
  • General Users: PrefPaint received near-perfect scores (4.3+) for realism and smoothness, while baselines struggled with blending and texture.
  • Usability: The web interface received high marks for adaptability and ease of use.

5. Significance and Impact

• Clinical Reliability: By integrating direct feedback from oncologists and gastroenterologists, PrefPaint ensures that synthetic training data adheres to real-world medical standards, reducing the risk of AI hallucinations in diagnostics.
• Democratization of AI Training: The system lowers the barrier to entry for medical institutions to fine-tune generative models, removing the computational bottleneck of reward model training.
• Future of Medical AI: This work demonstrates a viable path toward "Human-in-the-Loop" AI development in healthcare, where iterative expert feedback continuously refines model performance, leading to safer and more accurate diagnostic tools.