💬 NLP

CURE: Curriculum-guided Multi-task Training for Reliable Anatomy Grounded Report Generation

The paper introduces CURE, a data-efficient, error-aware curriculum learning framework that dynamically adjusts multi-task training to significantly enhance visual grounding, factual consistency, and hallucination reduction in automated medical report generation without requiring additional data.

Original authors: Pablo Messina, Andrés Villa, Juan León Alcázar, Karen Sánchez, Carlos Hinojosa, Denis Parra, Álvaro Soto, Bernard Ghanem

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pablo Messina, Andrés Villa, Juan León Alcázar, Karen Sánchez, Carlos Hinojosa, Denis Parra, Álvaro Soto, Bernard Ghanem

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are training a very smart, but slightly clumsy, medical student to write reports based on X-ray images. This student (a type of AI called a Vision-Language Model) is great at looking at pictures and writing words, but they have a bad habit: they often make things up.

If you show them an X-ray of a healthy chest, they might confidently write, "I see a broken bone," even though the bone is perfectly fine. In the medical world, this is called a "hallucination," and it's dangerous because it could lead to unnecessary panic or treatment.

The paper introduces a new training method called CURE (Curriculum-guided Multi-task Training for Reliable Anatomy Grounded Report Generation) to fix this. Here is how it works, using simple analogies:

1. The Problem: The "Guessing" Student

Current AI models are like students who memorize that "X-rays usually show broken bones." So, when they look at an X-ray, they immediately jump to the conclusion that there is a break, even if there isn't one. They also struggle to point exactly where on the image they see the problem. They might say, "There is a fracture," but if you ask them to point to it on the picture, their finger might be waving around in the wrong spot.

2. The Solution: The "CURE" Training Plan

The authors didn't just give the student more textbooks (more data). Instead, they changed how the student learns. They created a special training schedule, or "curriculum," that acts like a strict but helpful tutor.

Analogy: The "Focus on Weakness" Tutor
Imagine a gym coach who watches an athlete train.

Old Method: The coach makes the athlete run the same 10 laps every day, regardless of how they are doing.
CURE Method: The coach watches the athlete closely. If the athlete trips on the left turn, the coach says, "Okay, we are going to run 20 laps focusing only on the left turn today." If the athlete is great at sprinting but bad at jumping, the coach adjusts the workout to spend more time on jumping.

CURE does this for the AI. It constantly checks the AI's performance. If the AI is bad at describing the "left lung" or good at finding "broken bones" but bad at finding "pneumonia," CURE automatically adjusts the training to feed the AI more examples of the things it is struggling with. It forces the AI to stop guessing and start paying attention to the specific details it keeps missing.

3. The "Grounding" Lesson: Pointing to the Evidence

A major part of CURE is teaching the AI to "ground" its words.

Without CURE: The AI says, "I see a fracture." (But it might be lying, and you don't know where it is).
With CURE: The AI must say, "I see a fracture [points to specific spot on the image]."

The training forces the AI to link every single sentence it writes to a specific box drawn on the X-ray. If the AI can't find a spot on the image to match its words, it learns that it shouldn't say those words. This stops it from making up diseases that aren't there.

4. The Results: A More Honest Doctor

The paper tested this new training method against the current best AI models (like MAIRA-2) using public chest X-ray datasets. Here is what they found:

Less Lying: The AI made up (hallucinated) abnormal findings 18.6% less often.
Better Pointing: The AI got much better at pointing to the exact right spot on the X-ray (improving accuracy by a significant margin).
Better Reports: The written reports were more accurate and matched what human doctors would say.
No Extra Data Needed: The best part is that they didn't need to find new, secret medical records to train the AI. They just used the public data they already had but taught the AI in a smarter way.

Summary

Think of CURE as a new teaching style for medical AI. Instead of letting the AI guess and hope for the best, it acts like a strict tutor that says, "You keep getting the left collarbone wrong, so we are going to practice that until you get it right, and you must point to it on the picture every time you speak."

The result is an AI that is less likely to invent fake diseases and much better at showing exactly where it sees a problem, making it a more reliable tool for doctors.

Technical Summary: CURE

Problem Statement

Medical Vision-Language Models (VLMs) have demonstrated potential in automating radiology report generation, yet they face critical limitations regarding visual grounding and factual consistency. Existing state-of-the-art models, such as MAIRA-2, often misalign textual findings with visual evidence, leading to "hallucinations" where the model generates clinically incorrect descriptions or detects pathologies that are not present in the image. This lack of reliability hinders clinical adoption, as models must produce factually accurate outputs grounded in specific image regions to ensure trustworthiness in real-world workflows. Furthermore, standard fine-tuning strategies often fail to transfer visual grounding capabilities to models that lack them natively, and existing multi-task learning approaches struggle with imbalanced data distributions, causing models to overfit frequent patterns while neglecting rare but clinically significant findings.

Methodology

The authors propose CURE (CUrrriculum-guided Multi-task Training for REliable Anatomy Grounded Report Generation), a framework designed to enhance the reliability and visual grounding of medical VLMs without requiring additional training data. CURE operates on a base model (MedGemma-4B-IT) and reformulates the training pipeline through three core components:

1. Fine-Grained Task Formulation

CURE unifies diverse supervision signals (bounding boxes, phrases, anatomical labels, and descriptive sentences) from public datasets (MIMIC-CXR/Chest ImaGenome, MS-CXR, PadChest-GR) into a consistent multi-task instructional format. The framework decomposes tasks into:

Phrase Grounding (PG): Linking report phrases to bounding boxes.
Grounded Report Generation (GRG): Generating full reports with explicit coordinate links.
Anatomy-Grounded Report Generation (AGRG): A novel, detailed task derived from Chest ImaGenome scene graphs. It splits the generation process into three subtasks for specific anatomical locations: Locate (spatial localization), Describe (contextual description), and Locate and Describe (simultaneous execution). This decomposition allows the model to learn spatial and semantic alignment independently and jointly.

2. Error-Aware Curriculum Learning

To address data imbalance (both inter-dataset and intra-dataset) and prevent the model from overfitting to easy samples, CURE employs a dynamic curriculum learning strategy.

Dynamic Sampling: Instead of static sampling proportional to dataset size, CURE adjusts sampling probabilities based on the model's current performance.
Two-Level Granularity:
- Inter-Dataset: Re-weights the contribution of different data sources (e.g., MS-CXR vs. Chest ImaGenome) based on aggregate performance scores.
- Intra-Dataset: Re-weights specific anatomical regions or semantic classes within a dataset based on their per-class error rates.
Performance Metric: Sampling weights are updated iteratively using a composite score ( $s_i$ ) combining Intersection over Union (IoU) for localization and CXRFEScore for textual fidelity. The error rate ( $e_i = 1 - s_i$ ) dictates the sampling probability for the next stage, forcing the model to focus on "harder" samples and underperforming categories.

3. Training Pipeline

The training process involves a pre-training stage on Chest ImaGenome to establish a grounding baseline, followed by a multi-task fine-tuning phase where the curriculum learning mechanism is active. The framework utilizes LoRA (Low-Rank Adaptation) for efficient parameter tuning and employs bounding-box-aware data augmentations (e.g., spatial transforms, CLAHE) to improve robustness.

Key Contributions

CURE Framework: The introduction of an error-aware curriculum learning framework that dynamically adjusts sampling distributions based on model performance, enhancing visual grounding without requiring additional data.
Effective Transfer: Demonstrating that the proposed training strategy successfully transfers grounding capabilities to medical VLMs (specifically MedGemma-4B-IT) that originally lack visual grounding mechanisms, allowing them to surpass current state-of-the-art models despite being trained on less data.
State-of-the-Art Performance: Achieving a new state-of-the-art in visually grounded report generation, specifically reducing hallucinations and improving spatial accuracy across multiple benchmarks.

Results

The paper evaluates CURE against strong baselines, including MAIRA-2 and the base MedGemma-4B-IT, across several tasks and datasets:

Visual Grounding: CURE improves phrase grounding accuracy by +0.35 IoU on the Chest ImaGenome dataset compared to MAIRA-2, effectively doubling localization performance. It also achieves higher Micro and Macro IoU scores on MS-CXR, PadChest-GR, and the zero-shot VinDr-CXR dataset.
Report Quality: The framework boosts report quality by +0.192 CXRFEScore and achieves the highest scores in CheXbert F1-Micro, cosine similarity, and CXRFEScore on the Anatomy-Grounded Report Generation task.
Hallucination Reduction: CURE reduces abnormal finding hallucinations by 18.6% across six key anatomical regions. Specifically, it halves the rate of contradictions (17.44% vs. 33.22% for MAIRA-2) and more than doubles entailment rates (39.50% vs. 15.94%). The reduction is particularly significant for bone structures like the clavicles, where CURE's hallucination rate drops to 1.00% compared to MAIRA-2's >59%.
Zero-Shot Generalization: The model demonstrates strong generalization capabilities, outperforming MAIRA-2 in grounding metrics on the unseen VinDr-CXR dataset, despite MAIRA-2's training on proprietary data.

Significance and Claims

The paper claims that CURE represents a significant step forward in making medical VLMs reliable and trustworthy for clinical deployment. By prioritizing underperforming samples and anatomical regions through an error-aware curriculum, the framework addresses the fundamental issue of hallucination and misalignment that plagues current models. The authors emphasize that CURE achieves these improvements without requiring additional data or model parameters, relying instead on a more efficient and adaptive use of existing public datasets. The work suggests that dynamic, performance-driven curriculum learning is a viable and necessary strategy for bridging the gap between general-purpose multi-modal models and the rigorous demands of medical imaging, ultimately fostering greater transparency and trust in AI-assisted diagnostics.