PaperlayerView exact revision on arXiv

Source-linked AI summary

ECHO: Efficient Chest X-ray Report Generation with One-step Block Diffusion

Lifeng Chen, Tianqi You, Hao Liu, Zhimin Bao, Jile Jiao, Xiao Han, Zhicai Ou, Tao Sun, Xiaofeng Mou, Xiaojie Jin, Yi Xu

arXiv:2604.09450v2cs.LGcs.AIeess.IV

TL;DR

Autoregressive CXR report generation is slow, while one-step diffusion can lose coherence because token-factorized denoisers miss joint dependencies. ECHO uses DCD and RAD to enable efficient one-step-per-block generation, reporting higher quality metrics and substantial inference speedups over autoregressive baselines.

  • Problem

    Autoregressive VLMs have high inference latency from sequential decoding, while one-step diffusion can degrade coherence because token-factorized denoisers miss joint dependencies.

  • Method

    ECHO combines RAD-based AR-to-block-diffusion adaptation with DCD distillation from on-policy diffusion trajectories to enable one-step-per-block inference.

  • Results

    64.33% and 60.58% improvements in RaTE and SemScore, respectively, accompany up to 8× theoretical and 5.1× practical inference speedups over autoregressive baselines.

  • Takeaways & Limitations

    ECHO demonstrates that high clinical accuracy and extreme decoding efficiency are simultaneously achievable in automated chest X-ray reporting.

Abstract

from arXiv · show

Chest X-ray report generation (CXR-RG) has the potential to substantially alleviate radiologists' workload. However, conventional autoregressive vision--language models (VLMs) suffer from high inference latency due to sequential token decoding. Diffusion-based models offer a promising alternative through parallel generation, but they still require multiple denoising iterations. Compressing multi-step denoising to a single step could further reduce latency, but often degrades textual coherence due to the mean-field bias introduced by token-factorized denoisers. To address this challenge, we propose \textbf{ECHO}, an efficient diffusion-based VLM (dVLM) for chest X-ray report generation. ECHO enables stable one-step-per-block inference via a novel Direct Conditional Distillation (DCD) framework, which mitigates the mean-field limitation by constructing unfactorized supervision from on-policy diffusion trajectories to encode joint token dependencies. In addition, we introduce a Response-Asymmetric Diffusion (RAD) training strategy that further improves training efficiency while maintaining model effectiveness. Extensive experiments demonstrate that ECHO surpasses state-of-the-art autoregressive methods, improving RaTE and SemScore by \textbf{64.33\%} and \textbf{60.58\%} respectively, while achieving up to \textbf{$8\times$} inference speedup with negligible degradation in clinical accuracy.

1 Introduction

ECHO targets fast, coherent chest X-ray report generation by replacing sequential or multi-step decoding with one-step-per-block diffusion. Its DCD and RAD strategies support this efficiency while preserving clinical accuracy and improving reported quality metrics.

  • CXR report generation could reduce radiologists’ workload, but autoregressive VLMs incur high latency from sequential decoding.
  • One-step diffusion can maximize throughput but suffers incoherent outputs because token-factorized denoisers cannot represent joint token dependencies.
  • DCD constructs unfactorized supervision from on-policy teacher trajectories conditioned on committed high-confidence context.
  • ECHO converts an enhanced autoregressive CXR model into block diffusion with RAD, then distills it into a one-step counterpart using DCD.
  • 64.33% and 60.58% improvements in RaTE and SemScore, respectively, accompany up to 8× theoretical and 5.1× practical inference speedups over autoregressive baselines.
  • DCD reaches up to 390% speedup over the corresponding multi-step baseline at block size L=8 with marginal quality degradation, while RAD reduces theoretical training FLOPs by 72.3%.

2 Related Works

Related work develops parallel diffusion-based generation and acceleration through inference optimization and distillation. ECHO addresses the remaining throughput gap by enabling one-step denoising with unfactorized supervision.

  • Discrete diffusion language models progressively mask tokens and recover them through a learned reverse process over a discrete vocabulary.
  • dVLMs extend discrete diffusion to image-conditioned parallel generation through visual encoder alignment and instruction fine-tuning.
  • Existing acceleration methods reduce computation per step or distill multi-step denoising, but retain token-factorized prediction targets.
  • ECHO’s DCD supports one-step denoising for dLLMs and dVLMs, applied here to block diffusion.

3 Preliminaries

The preliminaries explain that dLLMs use token-factorized reverse predictions, which cannot represent cross-position dependencies. This mean-field bias is smallest with few masked tokens and largest when the sequence is fully masked, motivating multi-step sampling.

  • dLLMs define a forward process that progressively masks tokens and a reverse process that recovers them.
  • Mean-field parameterization: Mean-field parameterization approximates the joint posterior with independent per-token factors because the full joint requires exponentially many parameters.
  • Mean-field parameterization: Per-position cross-entropy training matches each token’s conditional marginal but cannot capture cross-positional correlations in the true posterior.
  • Mean-field parameterization: Mean-field bias quantifies irreducible joint dependence structure that no factorized distribution can represent.
  • Why multi-step sampling mitigates the bias: Mean-field bias grows monotonically with masked-token count, reaching its maximum when all tokens are masked.

4 Methodology

ECHO converts an autoregressive radiology VLM into a one-step-per-block diffusion model through RAD and DCD, while addressing training efficiency, joint token dependencies, termination, and inference overhead.

  • Pipeline overview: ECHO uses three successive stages: continued pre-training creates ECHOAR, RAD creates ECHOBase, and DCD produces the final ECHO model.The final model performs one-step-per-block decoding.
  • Response-Asymmetric Diffusion: RAD converts ECHOAR into a block diffusion model by duplicating only response tokens and letting noisy response blocks attend to context and prior decoded blocks.This avoids duplicating long vision-token sequences in a two-stage conversion.
  • Direct Conditional Distillation: DCD collects high-confidence distributions along the teacher’s multi-step trajectory and stitches them into a joint target that captures cross-token dependencies.This non-factorized supervision is designed to reduce the mean-field limitation of one-step prediction.
  • Direct Conditional Distillation: DCD aligns the student’s one-step block prediction with the teacher target using forward KL divergence and reweights positions according to their unmasking steps.Positions exposed to stronger mean-field bias receive stronger supervision.
  • Termination and hallucination: An additional cross-entropy loss targets <eos> tokens because their confidence is lower and more variable than content-token confidence, especially at larger block sizes.This addresses difficulty terminating generation and the associated repetition loops.
  • Inference optimization: Fused block KV caching removes the dedicated cache-update pass by combining the preceding block’s update with the current block’s denoising forward.The optimization targets the overhead that is especially costly for one-step decoding.

5 Experiments

Experiments evaluate ECHO across report quality, clinical fidelity, stability, efficiency, component ablations, and training-data effects. ECHO provides a strong quality–speed trade-off, while DCD, RAD, and report normalization each address distinct performance or efficiency requirements.

  • Evaluation setup: Experiments compare ECHO with proprietary, autoregressive, and diffusion-based models across linguistic, clinical, structural, and efficiency dimensions.The evaluation uses average metrics across MIMIC-CXR, CheXpert-Plus, and ReXGradient.
  • Overall performance: 8× theoretical speedup incurs only 3–7% quality degradation for ECHOblk8, while ECHOblk4 achieves 4× speedup with clinical-fidelity degradation within 3% on average.These configurations outperform the reported quality–speed trade-offs of T3D and dParallel.
  • DCD ablation: SW reduces PPL from 23.72 to 21.07 and raises CheXpert-Plus RaTEScore from 54.87 to 56.30, whereas RKL lowers MIMIC-CXR CIDEr from 3.65 to 3.48.The authors associate SW with stronger inter-token conditioning and prefer forward KL for preserving plausible clinical findings.
  • DCD ablation: CE supervision on <eos> produces the largest ablation gain, raising ROUGE-L from 52.44 to 56.14 and CIDEr from 3.65 to 4.05.PPL falls to 18.83, supporting reliable termination as important for avoiding repetition loops.
  • RAD adaptation: Quality reaches or surpasses the ECHOAR baseline within about 60 RAD steps, while throughput rises from 1.62 to 2.17, a 33.95% increase.Quality saturates at roughly 2.2% of the full RAD training data, whereas throughput continues improving later.
  • DCD ablation: DCD improves ECHOblk4 on CheXpert-Plus by 36% in ROUGE-L and 44% in CIDEr over native one-step decoding, with SemScore gains up to 29% on MIMIC-CXR.The improvements remain consistent across both block sizes and all three benchmarks.
  • Report normalization: Normalized reports raise CheXpert-Plus ROUGE-L from 23.82 to 56.89 and RaTEScore from 45.59 to 59.18 during Stage I.Without normalization, distillation later reduces CheXpert-Plus ROUGE-L to 18.79, while normalized training keeps SemScore above 45.0.

6 Conclusion

The conclusion presents ECHO as a discrete diffusion VLM that combines clinical accuracy with efficient decoding for automated chest X-ray reporting. It frames the approach as a foundation for broader medical multimodal generation and documents supplementary analyses supporting implementation and qualitative behavior.

  • Conclusion: ECHO demonstrates that high clinical accuracy and extreme decoding efficiency are simultaneously achievable in automated chest X-ray reporting.The conclusion positions this combination as the paper’s central outcome.
  • Conclusion: ECHO reduces training cost, eliminates multi-step denoising, and outperforms autoregressive and diffusion-based state-of-the-art methods across standard benchmarks.These claims summarize the paper’s reported system-level advantages.
  • Supplementary analyses: The supplementary material includes qualitative examples, experimental details, baseline implementations, detailed per-dataset results, and fused block KV-cache analysis.The KV-cache analysis formally shows preserved total FLOPs with half as many forward passes for one-step-per-block decoding.
  • Qualitative evidence: Qualitative comparisons show severe disorder and repetition in undistilled native one-step decoding, while DCD largely removes these artifacts in ECHOblk8 and ECHOblk4.Additional examples report accurate detection and clinical description of multiple abnormalities.

B.1 Data

The paper constructs a large, standardized, bilingual CXR report corpus from multiple public datasets through filtering, normalization, deduplication, translation, and supplemental multimodal augmentation.

  • Corpus construction: The unified corpus aggregates reports from MIMIC-CXR, CheXpert-Plus, ReXGradient, and IU-Xray.A five-stage preprocessing pipeline is applied across sources.
  • Preprocessing: Modality filtering removes non-CXR studies and samples with incomplete clinical descriptions using clinical entity extraction.This is the first stage of corpus preparation.
  • Preprocessing: Prompt-based rewriting standardizes terminology, enforces a structured Findings–Impression JSON format, and explicitly enumerates negative findings.The rewriting pipeline uses BaichuanM2-32B.
  • Preprocessing: Semantic deduplication leaves approximately 260k MIMIC-CXR, 250k CheXpert-Plus, 190k ReXGradient, and 30k IU-Xray reports.Near-duplicates are removed using Qwen3-Embedding-8B and a cosine-similarity threshold.
  • Preprocessing: Bilingual augmentation translates a random 50% of original reports from English into Chinese while preserving medical terminology and report structure.The pipeline also incorporates LLaVA-ReCap-558k to preserve general vision-language grounding.
  • Corpus construction: The resulting corpus is large-scale, standardized, and bilingual for CXR report generation.The corpus combines the preceding preprocessing stages into a unified training resource.

B.2 Training details

Training proceeds through autoregressive continual pretraining, diffusion adaptation, and distillation, with memory-conscious settings for high-resolution chest X-ray inputs.

  • Training configuration: All stages enforce a maximum image size of 2,250,000 pixels to balance high-resolution information retention with GPU memory consumption.Training is conducted on Alibaba PPUs.
  • Stage 1: AR continual pretraining: Stage 1 performs full-parameter autoregressive SFT on Lingshu using AdamW with a 1 × 10−5 learning rate and 0.01 weight decay.A linear warmup covers the first 0.03 epochs.
  • Stage 2: Diffusion adaptation: Stage 2 converts the model to diffusion using approximately 2.2% of Stage 1 data, freezing the vision encoder and projector while training only the LLM backbone.Neat-packing addresses the memory overhead of extended high-resolution vision-token sequences.
  • Stage 3: Distillation: Stage 3 distills the diffusion model using a randomly sampled 2.3% of the SFT dataset while keeping the vision encoder and projector frozen.This stage also uses neat-packing-based training.

B.3 Metric Details

The study evaluates report quality across clinical accuracy, semantic similarity, fluency, and decoding efficiency using established metrics and task-specific adaptations.

  • Evaluation framework: Five established metrics evaluate generated reports across quality and efficiency dimensions.Quality metrics include ROUGE-L, CIDEr, Modified RaTEScore, SemScore, and Perplexity; TPF and TPS assess decoding efficiency.
  • Quality metrics: ROUGE-L measures surface overlap through longest-common-subsequence structure and word-order similarity.
  • Quality metrics: CIDEr emphasizes rare clinical terms using cosine similarity between TF-IDF-weighted n-grams.
  • Quality metrics: Modified RaTEScore measures clinical entity-level accuracy using NER, medical embeddings, and a weighted F1 clinical-importance matrix.
  • Metric adaptation: Negative-related entity types receive zero importance weights because the normalized test set explicitly enumerates negative findings.This adaptation focuses evaluation on correctly identified positive pathologies.
  • Quality and efficiency metrics: SemScore measures semantic similarity with cosine similarity between sentence-transformer embeddings, while Perplexity measures generated-text fluency.TPF divides decoded tokens by model forward passes, and TPS divides generated tokens by elapsed decoding time.

C Implementation of baselines

The baseline implementation specifies decoding settings, standardization prompts, and several distillation strategies for controlled comparison.

  • Autoregressive baselines: Autoregressive baselines use greedy decoding, a 512-token limit, KV caching, and FlashAttention.
  • Diffusion baselines: LLaDA-MedV uses low-confidence remasking with response length L = 256, block length B = 64, and Z = 256 sampling steps.Its default fast_dllm configuration further accelerates inference.
  • Distillation baselines: Additional distillation baselines share the ECHO-Baseblk8 model and training data and use semi-autoregressive decoding aligned with BD3LM.They are grouped into trajectory-, distribution-, and optimization-based methods.
  • Report preprocessing: The standardization prompt separates Findings and Impression content and requires standardized reports in JSON format.The template addresses incomplete findings and misplaced interpretive statements, while preserving original findings and impressions.
  • Distillation baselines: Distribution-based distillation aligns predictions across noise levels with CE and KL losses, whereas optimization-based methods correct mean-field bias using teacher pseudo-labels.

D Detailed results

Detailed per-dataset tables compare ECHO with proprietary general models, autoregressive medical models, and diffusion methods across Chinese and English evaluations.

  • Per-dataset results: ECHOBase achieves state-of-the-art performance across all datasets, while ECHO maintains performance with minimal quality degradation and maximum inference speedup.

E Fused Block KV Cache Analysis

Fused Block KV Cache combines deferred cache updates with subsequent denoising, reducing one-step-per-block decoding passes without increasing FLOPs.

  • FLOPs equivalence: Fused Block KV Cache preserves total FLOPs while halving the number of forward passes, directly lowering inference latency.The equality holds for every n ≥ 1.
  • Vanilla Block KV Cache: Vanilla Block KV Cache performs separate denoise and KV-update passes for each block, yielding 2N forward passes.
  • Fused Block KV Cache: Fused Block KV Cache defers block 0’s update and merges each later update with the next block’s denoising, yielding N forward passes.
  • FLOPs equivalence: The fused operation preserves the corresponding attention contexts for decoded and masked tokens under the block-causal mask.
Loading 2604.09450v2…