💻 computer science

GRCD: Grounded Region Change Detection for Multi-Finding Chest X-Ray Pairs

This paper introduces GRCD, a novel framework that generates spatially grounded, multi-finding chest X-ray reports with accurate temporal change detection by leveraging a rigorously cleaned dataset and a Region-Guided Change Token module with a dual-pathway integration strategy.

Original authors: OFM Riaz Rahman Aranya, Peyman Najafirad, Kevin Desai

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: OFM Riaz Rahman Aranya, Peyman Najafirad, Kevin Desai

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 a doctor looking at two X-rays of a patient's chest: one taken today and one taken a year ago. Their job is to write a report that says, "The pneumonia in the left lung has gotten better," "The heart size is the same," and "There is a new shadow on the right side."

Doing this manually is hard work. The doctor has to spot the changes, remember where they are on the body, and write it all down. This paper introduces a new AI system called GRCD (Grounded Region Change Detection) that tries to do this job automatically.

Here is how the paper explains GRCD, using simple analogies:

1. The Problem: The "One-Note" Robot

Previous AI tools had two main flaws:

The "Snapshot" AI: It could look at just one X-ray and describe what it saw, but it couldn't compare it to an old one to see what changed.
The "Time-Travel" AI: It could compare two X-rays over time, but it couldn't point to where on the image the change happened. It was like a narrator describing a movie without showing you the screen.
The "Single-Story" AI (TRACE): There was one tool that could do both (compare time and point to spots), but it was limited. It could only describe one thing at a time. If a patient had three different problems changing in three different ways, this tool would get confused and just repeat the same thing over and over.

GRCD is the first tool that can look at two X-rays, find multiple different problems, say exactly how each one changed (better, worse, or the same), and draw a box around each specific spot on the image.

2. The Dirty Data: Cleaning the "Recipe Book"

Before building the robot, the authors had to fix the "recipe book" (the dataset) they were using to teach it. They found two major errors in the existing data:

The "Double-Exposure" Mistake: Sometimes, a radiographer takes two pictures of the same patient in the same minute to fix a blurry image. The old system thought these were two different days and tried to compare them. Since the pictures were identical, the AI learned nothing. The authors filtered these out.
The "Silent Ghost" Mistake: The old data relied on computer text to guess if a problem was "new" or "gone." But sometimes, a problem would just disappear from the text without a note. The computer would think it was "resolved" when it was actually still there, or vice versa. The authors created a three-step "detective" process to verify every single change, throwing out any data that looked suspicious. This resulted in a much cleaner, smaller, but much smarter dataset.

3. The Brain: The "Dual-Pathway" Strategy

The core of GRCD is a special module called RGCT (Region-Guided Change Token). Think of the AI's brain as having two ways to learn about the X-rays:

Pathway A (The "Headline" Approach): The AI takes the six most important body parts (like the heart and lungs) and puts a "sticky note" with a summary of their changes right at the very top of its reading list. This ensures the AI never forgets the big picture.
Pathway B (The "Deep Dive" Approach): As the AI writes its report, it constantly has the option to peek at a detailed list of 25 different body regions. It uses a "gate" to decide which details are important to look at right now.

The paper found that using both methods together works best. If you only use the "Headline," the AI gets lost in the details. If you only use the "Deep Dive," the AI forgets the main context. Together, they let the AI write a complex report with many different findings, each with its own location and status.

4. The Results: Better Than the Competition

When tested on a large set of real X-ray pairs:

Writing Quality: GRCD wrote reports that sounded much more natural and accurate than previous models.
Clinical Accuracy: It correctly identified medical conditions and their changes better than the competition.
Pointing Accuracy: It drew boxes around the problems with high precision.
The "Multi-Finding" Win: Unlike the previous best tool (TRACE), which collapsed into repeating the same sentence when faced with multiple problems, GRCD successfully handled reports with many different findings changing in different directions.

5. The Takeaway

The authors emphasize that quality is better than quantity. By spending time cleaning the data and removing the "noisy" or confusing examples, they built a smarter model than if they had just used a massive amount of messy data.

In short, GRCD is a new AI that acts like a careful radiologist's assistant: it compares today's X-ray to yesterday's, spots multiple changes, knows exactly where they are on the body, and writes a clear, structured report about what improved, what got worse, and what stayed the same.

Technical Summary: GRCD – Grounded Region Change Detection for Multi-Finding Chest X-Ray Pairs

Problem Statement
Radiologists routinely compare current and prior chest X-rays (CXRs) to track disease progression, generating follow-up reports that describe multiple findings, each localized to an anatomical region and annotated with a temporal change status (e.g., worsened, improved, stable). Existing automated approaches fail to address this complex, multi-finding setting holistically. Single-image methods generate fluent reports but lack temporal context. Temporal methods compare image pairs but fail to spatially ground their outputs. Grounded methods localize findings but operate on single time points. While TRACE [2] combines temporal reasoning with spatial grounding, it is restricted to single-finding descriptions. Real-world follow-up reports, however, contain multiple co-occurring findings with mixed change directions. Current models without explicit region-level encoding tend to collapse into repetitive, single-label outputs when faced with this multi-finding task.

Methodology
The authors propose GRCD (Grounded Region Change Detection), a framework designed to generate structured, multi-finding temporal reports with per-finding spatial grounding. The architecture consists of three primary components:

Data Construction and Cleaning: The authors constructed a benchmark dataset from MIMIC-CXR v2.1.0 and Chest ImaGenome v1.0. They identified and corrected two systematic labeling errors in the source annotations:
- Same-Study AP Retakes: Removing pairs where multiple images from a single exam (e.g., AP retakes) were incorrectly paired as prior/current comparisons.
- Silent NLP Absence: Addressing the issue where NLP-derived cues fail to label findings that appear or disappear without explicit textual markers.
- Three-Tier Filtering: Findings were classified into three tiers based on confidence. Tier 1 (direct NLP cues) and Tier 2 (recovered via keywords or temporal interpolation) were retained. Tier 3 (no evidence) pairs were discarded. This rigorous cleaning reduced the dataset from a naive 179,365 pairs to a high-confidence 40,250 pairs.
Region-Guided Change Token (RGCT) Module:
- A frozen BioViL-T temporal vision encoder processes the prior ( $I_p$ ) and current ( $I_c$ ) images to produce feature maps.
- Features are cropped from 25 anatomical regions defined by Chest ImaGenome bounding boxes.
- For each region, the module concatenates prior features, current features, and their difference. These descriptors pass through an inter-region self-attention block to allow information flow across regions.
- The refined descriptors are projected into the LLM hidden space to form change tokens. Auxiliary heads predict change labels, finding presence, and bounding boxes during training.
Dual-Pathway Integration:
- Pathway A (Prepended Tokens): Six tokens corresponding to key anatomical regions (lungs, heart, mediastinum, costophrenic angles) are prepended to the LLM input sequence. This provides a persistent, coarse-grained spatial signal.
- Pathway B (Gated Cross-Attention): All 25 change tokens are injected via gated cross-attention blocks at every fourth decoder layer. This allows the decoder to selectively query all regions during generation.
- The model is trained autoregressively to generate structured reports containing findings, change labels, and inline bounding boxes.

Key Contributions

First Multi-Finding Temporal Model: GRCD is the first model to generate multi-finding temporal reports with per-finding spatial grounding and distinct change labels.
Rigorous Dataset Cleaning: A three-tier pipeline that identifies and corrects systematic errors in Chest ImaGenome temporal pairs, resulting in a 40,250-pair benchmark.
Dual-Pathway Architecture: The introduction of the RGCT module with a dual-pathway integration strategy (prepended tokens + gated cross-attention), demonstrated to be necessary for successful multi-finding generation.

Experimental Results
Evaluated on a 7,967-sample multi-finding test set, GRCD outperformed existing baselines (BioViL-T classifier, nearest-neighbor retrieval, and re-evaluated TRACE) across all metrics:

Natural Language Generation: BLEU-4 improved to 0.398 (vs. 0.078 for TRACE and 0.273 for retrieval).
Clinical Accuracy: CheXbert F1 reached 0.567 (vs. 0.175 for TRACE) and RadGraph F1 reached 0.473 (vs. 0.222 for TRACE).
Change Detection: Achieved 53.3% report-level accuracy and 50.0% balanced accuracy, significantly outperforming the BioViL-T classifier (46.6%) and TRACE (39.8%).
Grounding: Achieved comparable grounding accuracy to TRACE (61.4% IoU > 0.5) but covered more samples (4,161 vs. 3,552).

Ablation studies confirmed that removing the RGCT module resulted in near-zero performance, while the dual-pathway design (Row D) outperformed either integration strategy (prepended-only or cross-attention-only) in isolation on text and clinical metrics. Furthermore, training on the larger, noisier uncleaned dataset degraded performance by up to 15.6%, validating the necessity of strict data cleaning.

Significance and Claims
The paper claims that GRCD addresses a critical gap in automated radiology reporting by enabling the generation of clinically realistic, multi-finding reports with spatial grounding. The authors emphasize that data quality is paramount over quantity for temporal tasks built on silver-standard annotations, as noisy labels significantly degrade model performance. They position GRCD as a decision-support tool intended to assist radiologists by reducing dictation time and flagging missed changes, rather than replacing human interpretation. The work provides a cleaned dataset and a dual-pathway architecture as building blocks for future research in temporal medical image understanding.