EMRRG: Efficient Radiology Report Generation

• EMRRG is a radiology report generation framework that efficiently fine-tunes a pre-trained Mamba-based vision encoder with a hybrid LLM decoder to produce detailed diagnostic reports.
• It leverages partial LoRA adaptation and gated cross-attention to fuse visual and textual features, achieving competitive performance on datasets like IU X-ray, MIMIC-CXR, and CheXpert Plus.
• EMRRG demonstrates state-of-the-art efficiency with reduced trainable parameters and lower runtime compared to conventional methods, making it a promising tool in medical AI.

Searching arXiv for the cited EMRRG paper and closely related context papers. EMRRG denotes “Efficient Fine-Tuning Pre-trained X-ray Mamba Networks for Radiology Report Generation,” a framework for chest X-ray medical report generation that couples a pre-trained Mamba-based vision encoder with a LLM decoder augmented by a hybrid cross-attention design (Zhang et al., 19 Oct 2025). In the supplied literature, the acronym is also adjacent to an unrelated construct: the “Emissions Reporting Maturity Model” (ERMM), which is presented as the most plausible match when “EMRRG” is used in the context of emissions-reporting governance, but that paper explicitly does not define or use the acronym EMRRG (Xavier et al., 2023). In current arXiv usage reflected by these sources, EMRRG therefore most directly refers to the radiology report generation framework, while ERMM is a separate maturity-model framework for city-level greenhouse-gas reporting.

1. Nomenclature and scope

The term EMRRG is explicitly defined in the radiology paper as a framework for generating free-text radiology reports from chest X-ray images by efficiently fine-tuning a pre-trained Mamba-based vision backbone and combining it with a LLM equipped with a specialized hybrid decoder (Zhang et al., 19 Oct 2025). The task setting is X-ray Medical Report Generation (MRG): given a chest X-ray image $I$, the model generates a report $Y = (y_1,\dots,y_T)$ in natural language, typically focusing on the “Findings” and “Impression” sections. The training objective is presented as autoregressive conditional likelihood maximization with a fixed instruction prompt, “Generate a comprehensive and detailed diagnosis report for this chest X-ray image.” (Zhang et al., 19 Oct 2025)

A distinct source in the provided corpus concerns climate-governance methodology rather than medical imaging. That paper consistently uses ERMM for “Emissions Reporting Maturity Model” and PIDP for “Performance Indicator Development Process,” and states that EMRRG is not mentioned explicitly (Xavier et al., 2023). This suggests that any encyclopedia treatment of EMRRG must distinguish the radiology-report-generation framework from the ERMM maturity model, despite the superficial orthographic similarity.

For the radiology usage, EMRRG is situated within multimodal medical AI, especially the subfield of automated report generation from chest radiographs. The central motivation is that existing MRG systems predominantly rely on LLMs, make limited use of pre-trained vision foundation models and advanced fine-tuning strategies, and rarely explore non-Transformer architectures such as Mamba in this application domain (Zhang et al., 19 Oct 2025).

2. Problem formulation and design rationale

EMRRG addresses the task of mapping a chest X-ray image to a textual diagnostic report under an end-to-end trainable encoder-decoder architecture (Zhang et al., 19 Oct 2025). The paper characterizes prior limitations along three axes: limited use of pre-trained vision foundation models with careful fine-tuning, naive fine-tuning strategies that either avoid adaptation or apply vanilla LoRA everywhere, and under-explored non-Transformer vision architectures for medical report generation (Zhang et al., 19 Oct 2025).

The model objective is written as

$\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$

with the prompt fixed in the experiments as the sentence quoted above (Zhang et al., 19 Oct 2025). No additional alignment or contrastive objectives are introduced; the reported training signal is the standard autoregressive language modeling loss.

The design rationale is threefold. First, EMRRG leverages a strong pre-trained Mamba-based vision encoder. Second, it adapts that encoder with partial LoRA specialized to Mamba internals rather than uniform full-rank or uniform LoRA adaptation. Third, it modifies the language-model decoder with hybrid decoder layers that compute self-attention over text and cross-attention from text to visual tokens in parallel, then fuse them through a gated residual mechanism (Zhang et al., 19 Oct 2025). This combination is intended to preserve parameter efficiency while improving cross-modal fusion.

A plausible implication is that EMRRG should be understood not merely as a new backbone selection, but as a coordinated PEFT-and-fusion strategy spanning both visual and linguistic subsystems. That interpretation is directly consistent with the paper’s emphasis on Efficient-Tuning SSM (ET-SSM), Partial LoRA, and the hybrid decoder as the key ingredients of the framework (Zhang et al., 19 Oct 2025).

3. Architecture: vision branch, language branch, and multimodal fusion

EMRRG is described as a two-branch encoder-decoder architecture composed of a vision branch and a language branch (Zhang et al., 19 Oct 2025). The input image is $I \in \mathbb{R}^{192\times 192 \times 3}$ and is divided into non-overlapping $16\times 16$ patches,

$$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$

with $N = 144$ for the specified image size (Zhang et al., 19 Oct 2025). A convolution with kernel size $16\times 16$ projects each patch into a 1024-dimensional token,

$T_i \in \mathbb{R}^{1024},\quad i = 1,\dots,N.$

These tokens are processed by a pre-trained Vision Mamba backbone, described as a ViM / MambaXray-VL variant based on state space models (Zhang et al., 19 Oct 2025).

The visual output is denoted $\mathbf{V}_s \in \mathbb{R}^{m\times d}$, with $Y = (y_1,\dots,y_T)$0, and is passed to a LLM that acts primarily as a text decoder (Zhang et al., 19 Oct 2025). The specific LLM varies by dataset: Qwen-1.5-1.8B is used on IU-Xray, while Llama2-7B is used on MIMIC-CXR and CheXpert Plus (Zhang et al., 19 Oct 2025).

Within the language branch, some standard decoder blocks are replaced by hybrid decoder layers. These layers run standard self-attention over text tokens and, in parallel, cross-attention between text queries and visual tokens; the outputs are then fused by a gated residual connection (Zhang et al., 19 Oct 2025). Four such hybrid layers are inserted at positions $Y = (y_1,\dots,y_T)$1 in the decoder stack (Zhang et al., 19 Oct 2025). This distributed insertion across depth is presented as the mechanism by which visual conditioning is propagated throughout generation rather than confined to a single fusion stage.

The framework is trained end-to-end with PEFT, with most backbone parameters frozen. The trainable subset comprises LoRA modules in the SSM and LLM together with the hybrid decoder components (Zhang et al., 19 Oct 2025). This operational choice is central to EMRRG’s identity: efficient adaptation is treated as a first-class architectural constraint rather than an implementation afterthought.

4. Mamba backbone and parameter-efficient adaptation

The vision encoder is based on Mamba, described in the paper as a selective state space model with a generic discrete update

$Y = (y_1,\dots,y_T)$2

where $Y = (y_1,\dots,y_T)$3 is the hidden state, $Y = (y_1,\dots,y_T)$4 is the input, and $Y = (y_1,\dots,y_T)$5 is the output (Zhang et al., 19 Oct 2025). The paper emphasizes that Mamba employs data-dependent, selective state updates and scan operations with linear complexity in sequence length, $Y = (y_1,\dots,y_T)$6, in contrast to the $Y = (y_1,\dots,y_T)$7 complexity associated with Transformer attention (Zhang et al., 19 Oct 2025).

Within the Mamba block, the paper identifies intermediate variables $Y = (y_1,\dots,y_T)$8, $Y = (y_1,\dots,y_T)$9, $\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$0, $\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$1, and $\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$2 as distinct internal components relevant to fine-tuning (Zhang et al., 19 Oct 2025). Rather than attaching LoRA to all such linear layers, EMRRG adopts the Partial LoRA strategy denoted $\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$3, in which LoRA is attached only to a subset of the weights associated with the intermediate feature $\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$4 (Zhang et al., 19 Oct 2025). The paper states that this is “partial” both because only particular modules are adapted and because LoRA is applied to part of the weight structure rather than the entire matrix.

The general LoRA formulation is given as

$\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$5

where $\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$6 and $\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$7, with low rank $\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$8 (Zhang et al., 19 Oct 2025). For the Mamba adaptation used in EMRRG, the paper writes conceptually

$\mathcal{L}_{\text{NLL} = -\sum_{i=1}^{T} \log p_\theta\!\big(y_i \mid \text{Prompt}, [y_1,\dots,y_{i-1}]\big),$9

with rank $I \in \mathbb{R}^{192\times 192 \times 3}$0 in the reported experiments (Zhang et al., 19 Oct 2025). Other Mamba parameters are frozen by default, except for standard LoRA on the input projection layer (Zhang et al., 19 Oct 2025).

This selective adaptation is presented as Efficient-Tuning SSM (ET-SSM). The paper’s ablations report that, among partial adaptations over $I \in \mathbb{R}^{192\times 192 \times 3}$1, $I \in \mathbb{R}^{192\times 192 \times 3}$2, $I \in \mathbb{R}^{192\times 192 \times 3}$3, $I \in \mathbb{R}^{192\times 192 \times 3}$4, and $I \in \mathbb{R}^{192\times 192 \times 3}$5, $I \in \mathbb{R}^{192\times 192 \times 3}$6 achieves the best overall balance on IU-Xray, with BLEU-1 $I \in \mathbb{R}^{192\times 192 \times 3}$7, BLEU-2 $I \in \mathbb{R}^{192\times 192 \times 3}$8, BLEU-4 $I \in \mathbb{R}^{192\times 192 \times 3}$9, ROUGE-L $16\times 16$0, and METEOR $16\times 16$1 (Zhang et al., 19 Oct 2025). The same ablation indicates that under fixed LLM tuning, $16\times 16$2 matches or even surpasses full fine-tuning on several metrics while updating only about $16\times 16$3 of parameters (Zhang et al., 19 Oct 2025). This suggests that EMRRG’s PEFT design is not only computationally motivated but also empirically aligned with the representational structure of Mamba.

5. Hybrid decoder and gated cross-attention

The hybrid decoder is the principal language-side innovation in EMRRG. A standard Transformer decoder layer is described as self-attention plus feed-forward transformation with residual and normalization structure; EMRRG augments this by adding a parallel cross-attention path from text to vision (Zhang et al., 19 Oct 2025).

Let $16\times 16$4 denote text hidden states and $16\times 16$5 denote visual features. The cross-attention output is written as

$16\times 16$6

where $16\times 16$7 are text queries and $16\times 16$8 map visual features into key and value spaces (Zhang et al., 19 Oct 2025). The paper explicitly states that self-attention and cross-attention are computed in parallel rather than interleaved, and that their outputs are fused through a dynamic gate.

The residual update is given as

$16\times 16$9

where $$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$0 is a token-wise, dimension-wise gate predicted from $$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$1 through a linear layer followed by $$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$2, and $$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$3 is a scalar warm-up gate initialized to $$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$4 and learned over time (Zhang et al., 19 Oct 2025). The paper explains the training behavior: initially, $$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$5, so the model behaves like a text-only LLM; as training proceeds, $$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$6 increases and the model gradually incorporates more visual information (Zhang et al., 19 Oct 2025).

The stated purpose of this two-level gating is to prevent cross-attention outputs from destabilizing a pre-trained LLM and to avoid noisy visual information overwhelming the text stream (Zhang et al., 19 Oct 2025). The hybrid layer thus combines linguistic coherence from self-attention, visual grounding from cross-attention, and selective, stable integration from gating (Zhang et al., 19 Oct 2025). In the component analysis, the paper reports that under fixed SSM settings, Llama2+HDL outperforms Llama2 or Llama2+LoRA alone on almost all metrics, which is offered as evidence for the value of the hybrid decoder (Zhang et al., 19 Oct 2025).

6. Empirical evaluation, datasets, and efficiency profile

EMRRG is evaluated on three benchmark datasets: IU X-ray, MIMIC-CXR, and CheXpert Plus (Zhang et al., 19 Oct 2025). IU X-ray is described as containing 7,470 images and 3,955 reports, with a 7:1:2 split under the same protocol as R2GenGPT (Zhang et al., 19 Oct 2025). MIMIC-CXR contains 377,110 images, 227,835 reports, and 65,379 patients, with train, validation, and test splits following R2GenGPT (Zhang et al., 19 Oct 2025). CheXpert Plus contains 223,228 chest X-rays and 187,711 reports, parsed into 11 sections, with “Findings” used as ground-truth text and a 7:1:2 split as in R2GenCSR (Zhang et al., 19 Oct 2025).

The paper uses BLEU-1 through BLEU-4, ROUGE-L, METEOR, and CIDEr as natural-language-generation metrics (Zhang et al., 19 Oct 2025). Clinical efficacy is evaluated through precision, recall, and F1 derived from automatic labelers, with the standard definitions

$$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$7

(Zhang et al., 19 Oct 2025).

On IU X-ray, EMRRG achieves BLEU-1 $$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$8, BLEU-2 $$P_i \in \mathbb{R}^{16\times 16\times 3},\quad i = 1,\dots,N,$$9, BLEU-3 $N = 144$0, BLEU-4 $N = 144$1, ROUGE-L $N = 144$2, METEOR $N = 144$3, and CIDEr $N = 144$4 (Zhang et al., 19 Oct 2025). On MIMIC-CXR, it achieves BLEU-1 $N = 144$5, BLEU-2 $N = 144$6, BLEU-3 $N = 144$7, BLEU-4 $N = 144$8, ROUGE-L $N = 144$9, METEOR $16\times 16$0, and CIDEr $16\times 16$1 (Zhang et al., 19 Oct 2025). On CheXpert Plus, it achieves BLEU-1 $16\times 16$2, BLEU-2 $16\times 16$3, BLEU-3 $16\times 16$4, BLEU-4 $16\times 16$5, ROUGE-L $16\times 16$6, METEOR $16\times 16$7, and CIDEr $16\times 16$8 (Zhang et al., 19 Oct 2025). The paper describes these CheXpert Plus results as new state of the art in lexical NLG metrics (Zhang et al., 19 Oct 2025).

Clinical-metric reporting further differentiates dataset behavior. On MIMIC-CXR, EMRRG attains precision $16\times 16$9, recall $T_i \in \mathbb{R}^{1024},\quad i = 1,\dots,N.$0, and F1 $T_i \in \mathbb{R}^{1024},\quad i = 1,\dots,N.$1, with F1 reported as second best in the compared set (Zhang et al., 19 Oct 2025). On CheXpert Plus, it attains precision $T_i \in \mathbb{R}^{1024},\quad i = 1,\dots,N.$2, recall $T_i \in \mathbb{R}^{1024},\quad i = 1,\dots,N.$3, and F1 $T_i \in \mathbb{R}^{1024},\quad i = 1,\dots,N.$4; precision is reported as the highest among earlier methods, which the paper interprets as suggesting fewer spurious findings (Zhang et al., 19 Oct 2025).

The efficiency profile is one of the framework’s strongest explicit claims. On CheXpert Plus, Table 8 reports runtime and trainable parameter counts, with EMRRG requiring 26.84 minutes and 1.32M trainable parameters (Zhang et al., 19 Oct 2025). This is contrasted in the paper with baselines such as R2GenGPT at 77.80 minutes and 90.90M tunable parameters, R2GenCSR at 31.20 minutes and 91.70M parameters, MambaXray-VL at 50.66 minutes and 57.31M parameters, and a ViT+Llama2 system at 10.82 minutes but 358.80M parameters (Zhang et al., 19 Oct 2025). The paper accordingly characterizes EMRRG as highly parameter- and compute-efficient.

7. Limitations, interpretation, and relation to ERMM

The EMRRG paper states several limitations. Experiments focus on hardware-friendly LLMs, specifically Qwen-1.8B and Llama2-7B, rather than larger models such as 13B or 70B variants (Zhang et al., 19 Oct 2025). The exploration of Mamba tuning is incomplete in the sense that not all internal components were fine-tuned or exhaustively ablated (Zhang et al., 19 Oct 2025). Performance is uneven across datasets: although EMRRG is especially strong on CheXpert Plus, it does not dominate every metric on IU-Xray and MIMIC-CXR, especially CIDEr (Zhang et al., 19 Oct 2025). The paper also notes that evaluation remains based on automatic NLG and label metrics rather than prospective clinical trials or real-world deployment studies (Zhang et al., 19 Oct 2025).

Future directions suggested in the paper include larger or more specialized LLMs for radiology, more fine-grained PEFT for Mamba, enhanced alignment objectives such as contrastive or retrieval-based losses, incorporation of RadGraph or GREEN for improved faithfulness, more sophisticated clinical evaluation, and possible integration into PACS or radiology workflow as a draft-report assistant (Zhang et al., 19 Oct 2025). A plausible implication is that EMRRG is best viewed as a modular foundation for efficient multimodal report generation rather than a finalized clinical system.

The separate ERMM literature is relevant primarily to terminology management. ERMM is a CMM-inspired maturity model for evaluating and classifying cities according to the quality and maturity of their emissions reporting processes, supported by PIDP and AI-based clustering methods such as WiSARD, ClusWiSARD, hierarchical clustering, and k-means (Xavier et al., 2023). That framework belongs to emissions-reporting governance rather than radiology, and the paper explicitly states that “EMRRG” is not defined or used there (Xavier et al., 2023). Thus, while both EMRRG and ERMM are AI-supported frameworks with process-oriented evaluation components, they address entirely different domains: one concerns chest X-ray report generation, the other city-level greenhouse-gas reporting maturity.

In summary, EMRRG in the strict acronymic sense refers to a radiology report generation framework built around a pre-trained Mamba vision encoder, Partial LoRA adaptation, and a hybrid LLM decoder with gated cross-attention (Zhang et al., 19 Oct 2025). Its significance lies in demonstrating that non-Transformer vision backbones and carefully localized PEFT can achieve competitive or state-of-the-art performance in medical report generation while maintaining a sharply reduced trainable-parameter budget (Zhang et al., 19 Oct 2025).