通过具有一致性嵌入的大语言模型实现端到端乳腺癌放射治疗计划制定|文献速递-最新论文分享

发布于:2025-07-03 ⋅ 阅读:(18) ⋅ 点赞:(0)

Title

题目

End-to-end breast cancer radiotherapy planning via LMMs with consistency  embedding

通过具有一致性嵌入的大语言模型实现端到端乳腺癌放射治疗计划制定

01

文献速递介绍

近年来,受大型语言模型(LLM)启发的新一代人工智能模型(即基础模型)的出现,标志着其与以往范式存在显著差异(Moor 等人,2023)。这些模型具有规模庞大、功能多样的特点,这源于它们在多样化数据上进行的自监督训练。目前,基础模型已能在多个领域实现最先进(SOTA)的性能,包括多模态推理、图文生成、图像 captioning 以及文本引导的图像分割等任务(Bubeck 等人,2023;Dai 等人,2024;Driess 等人,2023;Li 等人,2023b;Liu 等人,2024;Lai 等人,2024)。这些特性意味着人工智能融入医疗实践的方式可能发生范式转变——医疗实践本质上依赖多模态信息来制定全面的临床决策。此外,这也为克服目前 500 多种 FDA 批准的人工智能模型的局限性提供了机会,这些模型大多仅针对特定任务,且依赖单模态信息(Joshi 等人,2024)。具体而言,与这些单模态人工智能不同,结合基础模型的通用医疗人工智能能够全面理解临床工作流程,可处理多种医疗数据,包括影像模态、电子健康记录、实验室结果、基因组学甚至临床报告(Singhal 等人,2023;Rajpurkar 和 Lungren,2023;Wu 等人,2023b;Moor 等人,2023;Tu 等人,2024)。通过理解各类数据及其相互关系,多模态人工智能能提供患者数据的全面视图,从而促进更准确的诊断、个性化治疗方案的制定,并减少医疗差错。  ### 2.   本文聚焦放射肿瘤学领域,该领域中多模态整合至关重要,使其成为评估基础模型潜力的最重要临床领域之一。因此,我们在此介绍 RO-LMM——一种专为支持放射肿瘤学临床工作流程设计的大型多模态模型(LMM)原型。具体而言,本研究显著扩展了我们先前的相关工作 LLMSeg(Oh 等人,2024),后者侧重于多模态分割。更具体地说,RO-LMM 通过处理放射肿瘤学中更广泛的临床任务,扩大了 LLMSeg 的应用范围:(1)它能将海量患者病史和检查结果高效总结为简洁且信息丰富的临床记录;此外,它还能够(2)从临床专家视角提出合适的放射治疗策略,以及(3)在三维(3D)计算机断层扫描(CT)图像上勾勒出与所提放射治疗策略一致的放射靶区。RO-LMM 的这种多方面功能,在支持临床专业人员的专业工作方面展现出显著进展。 在训练 LLM 执行从放射治疗策略建议到靶区分割的一系列连续任务时,我们发现每个任务都存在误差累积的可能性,这可能导致端到端性能的显著下降。因此,本研究的另一重要贡献是借鉴并扩展了“噪声嵌入微调(NEFTune)”技术(Jain 等人,2024)——该技术在每个目标任务的训练过程中,向嵌入中注入均匀噪声。更具体地说,为进一步增强模型的适用性,我们开发了一种新颖的“一致性嵌入微调(CEFTune)”技术,通过添加正则化损失来确保模型对含噪输入和干净输入的预测一致性。此外,通过扩展到文本相关任务之外,我们将这些概念应用于 3D 分割任务,形成了新颖的“噪声嵌入分割(NESEG)”和“一致性嵌入分割(CESEG)”技术。这些进展防止了后续任务之间的误差传播,共同显著提升了端到端模型在内部和外部验证中的泛化能力。 作为概念验证研究,我们的 RO-LMM 框架被用于乳腺癌研究——乳腺癌是一种高发癌症,其放射治疗相对标准化,且仅需基于 CT 影像即可进行。 我们的贡献总结如下:   - 提出了一个全面的框架(称为 RO-LMM),其中 LMM 为乳腺癌放射治疗的广泛工作流程提供支持。据我们所知,该原型是首个支持放射肿瘤学全面工作流程的模型。   - 为防止在连续临床任务(如临床背景总结、放射治疗策略建议和基于计划的靶区分割)中出现潜在的误差累积,我们探索了噪声增强和一致性方法,并提出了新颖的训练方法(如 CEFTune、NESEG 和 CESEG),显著增强了方法的稳健性。   - 通过在乳腺癌患者真实临床数据的多种验证场景下进行实验,我们证明了 RO-LMM 的性能优于传统方法

Abatract

摘要

Recent advances in AI foundation models have significant potential for lightening the clinical workload by  mimicking the comprehensive and multi-faceted approaches used by medical professionals. In the field of  radiation oncology, the integration of multiple modalities holds great importance, so the opportunity of  foundational model is abundant. Inspired by this, here we present RO-LMM, a multi-purpose, comprehensive  large multimodal model (LMM) tailored for the field of radiation oncology. This model effectively manages  a series of tasks within the clinical workflow, including clinical context summarization, radiotherapy strategy  suggestion, and plan-guided target volume segmentation by leveraging the capabilities of LMM. In particular, to  perform consecutive clinical tasks without error accumulation, we present a novel Consistency Embedding FineTuning (CEFTune) technique, which boosts LMM’s robustness to noisy inputs while preserving the consistency  of handling clean inputs. We further extend this concept to LMM-driven segmentation framework, leading to a  novel Consistency Embedding Segmentation (CESEG) techniques. Experimental results including multi-center  validation confirm that our RO-LMM with CEFTune and CESEG results in promising performance for multiple  clinical tasks with generalization capabilities.

人工智能基础模型的最新进展具有巨大潜力,它能模仿医疗专业人员所采用的全面、多层面方法,从而减轻临床工作负担。在放射肿瘤学领域,多种模态的整合至关重要,因此基础模型的应用前景十分广阔。受此启发,我们提出了RO-LMM——一种专为放射肿瘤学领域设计的多用途、综合性大型多模态模型(LMM)。该模型借助大型多模态模型的能力,可有效处理临床工作流中的一系列任务,包括临床背景总结、放射治疗策略建议以及基于计划的靶区分割。 特别值得一提的是,为了在执行连续临床任务时避免误差累积,我们提出了一种新颖的一致性嵌入微调(CEFTune)技术,该技术能增强大型多模态模型对含噪输入的稳健性,同时保持处理干净输入时的一致性。我们进一步将这一概念扩展到由大型多模态模型驱动的分割框架中,形成了一种新颖的一致性嵌入分割(CESEG)技术。包括多中心验证在内的实验结果证实,我们结合了CEFTune和CESEG的RO-LMM在多项临床任务中表现出色,且具有良好的泛化能力。

Method

方法

In this section, we provide a detailed description of our proposed  approach designed for sequential text generation tasks, including summarization and suggestions, as well as text-driven image segmentation,  whose robustness is improved by consistency embedding finetuning.  The overall framework is illustrated in Fig. 2. 3.1. Consistency embedding fine-tuning for clinical LMM To realize the multi-purpose LMM with expertise in clinical report  summarization and radiotherapy strategy suggestion, we conduct instruction fine-tuning for LLaMA2 (Touvron et al., 2023). Considering  the nuanced differences in the intended objective of each task, we adopt  separate training strategies to acquire task-specific expertise, namely  RO-LMM-S (summary expert) and RO-LMM-P (plan expert). Specifically, we train a summary expert using collected raw clinical  report and summary notes. During inference, the summary expert  receives raw clinical reports, just like in the training scenario. However,  for the plan expert, there is a discrepancy between the training and  inference scenarios. In other words, we use the training set made up of  collected summary notes instead of generated notes from the summary  expert, mainly due to cost concerns and the inherent nature of our  framework, which generates output sequentially as illustrated in Fig.  1. However, at the inference phase, our model takes the generated  notes from the trained summary expert. To deal with the input domain  differences for the training and inference time, Noisy Embedding FineTuning (NEFTune) (Jain et al., 2024) which inject uniform noise into  embedding could be an effective naive solution to handle noisy inputs  in this task. However, a crucial consideration arises from the nature  of the generated notes, since some of them may lie closer to clean  inputs (collected notes) and the others deviate towards noisy inputs. To  address this, it is essential to train the model to handle both clean and  noisy inputs. To preserve the robustness facilitated by NEFTune while  enforcing consistency between the prediction given clean and noisy  inputs, we introduce Consistency Embedding Fine-Tuning (CEFTune),  resulting in RO-LMM-P++. 2 More details are as follows

在本节中,我们详细描述所提出的方法,该方法适用于序列文本生成任务(包括总结和建议)以及文本驱动的图像分割,其鲁棒性通过一致性嵌入微调得到提升。整体框架如图2所示。 3.1. 用于临床LMM的一致性嵌入微调 为实现具备临床报告总结和放射治疗策略建议专业能力的多用途LMM,我们对LLaMA2(Touvron等人,2023)进行指令微调。考虑到每项任务的预期目标存在细微差异,我们采用单独的训练策略来获取特定任务的专业知识,即RO-LMM-S(总结专家)和RO-LMM-P(计划专家)。 具体而言,我们使用收集到的原始临床报告和总结笔记来训练总结专家。在推理阶段,总结专家接收原始临床报告,与训练场景一致。然而,对于计划专家,训练场景和推理场景存在差异。也就是说,我们使用由收集到的总结笔记组成的训练集,而非来自总结专家生成的笔记,这主要是出于成本考虑以及我们框架的固有特性(如 图1所示,该框架会按顺序生成输出)。但在推理阶段,我们的模型会接收来自训练好的总结专家生成的笔记。 为处理训练和推理时的输入领域差异,在嵌入中注入均匀噪声的噪声嵌入微调(NEFTune)(Jain等人,2024)可能是处理该任务中噪声输入的一种有效的简单解决方案。然而,生成笔记的性质带来了一个关键问题,因为其中一些笔记可能更接近干净输入(收集到的笔记),而另一些则更偏向噪声输入。为解决这一问题,训练模型以同时处理干净输入和噪声输入至关重要。为在保持NEFTune所带来的鲁棒性的同时,确保对干净输入和噪声输入的预测一致性,我们引入了一致性嵌入微调(CEFTune),从而得到RO-LMM-P++。更多细节如下。 注:2 “+”、“++”分别表示采用NEFTune和CEFTune。

Conclusion

结论

In this work, we introduce RO-LMM, a multi-purpose, comprehensive foundation model tailored for radiation oncology. Addressing  limitations in current medical AI models confined to specific tasks, ROLMM demonstrates proficiency in diverse tasks encompassing overall  workflow of radiation oncology: clinical report summarization, radiotherapy strategy suggestion, and plan-guided 3D target volume segmentation. Another key contribution of this work is the introduction  of consistency technique into both text and segmentation task. Results  from multi-center cohort datasets confirm RO-LMM’s promising performance and noteworthy generalization capabilities across diverse tasks.  These findings mark a significant stride towards developing a versatile  AI model, hinting at the potential for a multi-purpose medical AI model  in radiation oncology

在本研究中,我们介绍了RO-LMM——一种专为放射肿瘤学设计的多用途、综合性基础模型。为解决当前医疗人工智能模型局限于特定任务的问题,RO-LMM在放射肿瘤学的整体工作流程中展现出对多种任务的处理能力,包括临床报告总结、放射治疗策略建议以及基于计划的三维靶区分割。本研究的另一关键贡献是将一致性技术引入文本任务和分割任务中。来自多中心队列数据集的结果证实,RO-LMM在各类任务中均表现出良好的性能和显著的泛化能力。这些发现标志着在开发多功能人工智能模型方面迈出了重要一步,也暗示了多用途医疗人工智能模型在放射肿瘤学领域的应用潜力。

Results

结果

5.1. Clinical report summarization We present the performance of our model on the clinical report  summarization task, along with confidence intervals for each method,  in Table 2. Our fine-tuned model of RO-LMM-S demonstrate significant improvements over the Defaults, providing consistent margins in  all metrics and confidence intervals. Notably, RO-LMM-S outperforms  ChatGPT with few-shot in-context learning. Moreover, we evaluate the generated summaries using expertisebased rubrics by two clinical experts and compare them to Defaults,  including ChatGPT and LLaMa-2. As shown in Table 3, our RO-LMM-S  model significantly outperforms all Defaults in both internal and external validations, thanks to its domain-specific knowledge. Additionally,  Pearson correlation (??) analysis reveals strong positive inter-clinician  correlations (> 0.85 and > 0.95 for internal and external validation,  respectively), confirming the reliability of our rubrics and the clinical  relevance of RO-LMM-S. Therefore, our RO-LMM-S provides practical  and meaningful summaries that can assist in the field of radiation  oncology.

5.1. 临床报告总结   我们在表2中呈现了模型在临床报告总结任务上的性能,以及每种方法的置信区间。经过微调的RO-LMM-S模型相较于传统方法(Defaults)有显著改进,在所有指标和置信区间上均保持稳定优势。值得注意的是,RO-LMM-S的性能优于采用少样本上下文学习的ChatGPT。   此外,我们通过两位临床专家基于专业评分标准对生成的总结进行了评估,并与包括ChatGPT和LLaMa-2在内的传统方法进行了对比。如表3所示,得益于其领域特定知识,我们的RO-LMM-S模型在内部和外部验证中均显著优于所有传统方法。此外,皮尔逊相关分析(??)显示,临床专家之间存在强正相关(内部验证>0.85,外部验证>0.95),这证实了我们评分标准的可靠性以及RO-LMM-S的临床相关性。因此,我们的RO-LMM-S能够提供实用且有意义的总结,可为放射肿瘤学领域提供辅助。

Figure

图片

Fig. 1. RO-LMM as an assistant large multimodal model (LMM) in the field of radiation oncology. The model seamlessly covers various tasks such as clinical report summarization,  radiation radiotherapy strategy suggestion, and 3D target volume segmentation.

图1. RO-LMM作为放射肿瘤学领域的辅助大型多模态模型(LMM)。该模型无缝涵盖多种任务,如临床报告总结、放射治疗策略建议以及三维靶区分割。

图片

Fig. 2. Schematics of RO-LMM training for three different tasks. (a) RO-LMM-S for clinical note summarization. (b) RO-LMM-P++ for radiotherapy strategy suggestion. (c)  RO-LMM-SEG++ for plan-guided target volume segmentation.

图2. RO-LMM针对三项不同任务的训练示意图。(a) 用于临床记录总结的RO-LMM-S;(b) 用于放射治疗策略建议的RO-LMM-P++;(c) 用于基于计划的靶区分割的RO-LMM-SEG++。

图片

Fig. 3. Schematics of RO-LMM-SEG++ for plan-guided 3D target volume segmentation task, which composed of (a) image module and (b) text module. These module outputs are  aligned through (c) multimodal alignment module.

图3. 用于基于计划的三维靶区分割任务的RO-LMM-SEG++示意图,该模型由(a)图像模块和(b)文本模块组成。这些模块的输出通过(c)多模态对齐模块进行对齐。

图片

Fig. 4. Qualitative comparison on 3D target volume segmentation task. Red arrows indicate errors.

图4. 三维靶区分割任务的定性对比。红色箭头指示错误之处。

Table

图片

Table 1 Training data details. CRS: Clinical Report Summarization. RSS: Radiotherapy Strategy  Suggestion. PTS: Plan-guided Target Segmentation. US: Ultrasound. Path: Pathology.

表1 训练数据详情   CRS:临床报告总结   RSS:放射治疗策略建议   PTS:基于计划的靶区分割   US:超声   Path:病理学

图片

Table 2 Quantitative comparison for clinical note summarization. Vanilla: the instruction fine tuning. CI: confidence interval.

表2 临床记录总结的定量对比   Vanilla:指令微调   CI:置信区间

图片

Table 3 Clinical expert analysis for report summarization. R#: each rubric, C#:each clinical  expert.

表3 报告总结的临床专家分析   R#:各项评分标准   C#:各位临床专家

图片

Table 4 Clinical expert analysis for radiotherapy strategy suggestion. R#: each rubric, C#: each clinical expert.

表4 放射治疗策略建议的临床专家分析   R#:各项评分标准   C#:各位临床专家

图片

Table 5 Comparison of 3D target volume segmentation performance

表5 三维靶区分割性能对比

图片

Table 6 Comparison of 3D target segmentation performance for overall and specific patient types

表6 针对整体及特定患者类型的三维靶区分割性能对比

图片

Table 7 Quantitative comparison results for our RO-LMM’s clinical report summarization and  radiotherapy strategy suggestion performance on the publicly available dataset.

表7 我们的RO-LMM在公开数据集上的临床报告总结和放射治疗策略建议性能的定量对比结果。

图片

Table 8 Ablation study on adopting separate expertise for each textual task against unified  strategy.

表8 针对每项文本任务采用单独专业知识与采用统一策略的消融研究对比

图片

Table 9 Ablation study on CESEG for target segmentation performance with input text variation.

表9 针对输入文本变化下靶区分割性能的CESEG消融研究

图片

Table 10 Component analysis of our proposed method on radiotherapy strategy suggestion..

表10 我们提出的方法在放射治疗策略建议方面的组件分析

图片

Table 11 Inference computational complexity. External validation (N

表11 推理计算复杂度。外部验证(N

图片

Table A.1 The proposed expertise-based rubrics for assessing the performance of clinical report  summarization.

表A.1 用于评估临床报告总结性能的基于专业知识的评分标准

图片

Table A.2 Score rubrics for radiotherapy strategy suggestion.

表A.2 放射治疗策略建议的评分标准