Search research articles

关于 JoVE

概览领导团队博客 JoVE 帮助中心

作者

出版流程编辑委员会范围与政策同行评审常见问题投稿

图书馆员

用户评价订阅访问资源图书馆顾问委员会常见问题

研究

JoVE Journal Methods Collections JoVE Encyclopedia of Experiments 存档

教育

JoVE Core JoVE Business JoVE Science Education JoVE Lab Manual 教师资源中心教师网站

使用条款与条件

相关概念视频

Vision

Vision

Vision is the result of light being detected and transduced into neural signals by the retina of the eye. This information is then further analyzed and interpreted by the brain. First, light enters the front of the eye and is focused by the cornea and lens onto the retina—a thin sheet of neural tissue lining the back of the eye. Because of refraction through the convex lens of the eye, images are projected onto the retina upside-down and reversed.

您也可能阅读

相关文章

通过共同作者、期刊和引用图与本文相关的文章。

排序

Same author

AI/ML-Assisted Detection of <i>HMGA2</i> RNA Isoforms in Prostate Cancer Patient Tissue.

International journal of molecular sciences·2026

Same author

A Multimodal Adaptive Inter-Region Attention-Guided Network for Brain Tumor Classification.

IEEE access : practical innovations, open solutions·2025

Same author

A Hybrid Learning-Architecture for Mental Disorder Detection Using Emotion Recognition.

IEEE access : practical innovations, open solutions·2024

Same author

Multi-Stage Classification of Retinal OCT Using Multi-Scale Ensemble Deep Architecture.

Bioengineering (Basel, Switzerland)·2023

Same author

Multi-Stage Classification-Based Deep Learning for Gleason System Grading Using Histopathological Images.

Cancers·2022

Same author

Predicting the Level of Respiratory Support in COVID-19 Patients Using Machine Learning.

Bioengineering (Basel, Switzerland)·2022

Search research articles

相关实验视频

Updated: Jan 7, 2026

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness

Published on: December 6, 2024

对用于放射学图像标题的低级调整视觉语言模型的实证评估.

Mahmudul Hoque¹, Raisa Nusrat Chowdhury¹, Md Rakibul Hasan²

¹Department of Computer Science, Morgan State University, Baltimore, MD 21251, USA.

Bioengineering (Basel, Switzerland)

|December 30, 2025

概括

大型视觉语言模型 (VLMs) 在医学成像任务中通常优于较小的模型. 然而,有针对性的适应策略和建筑设计允许一些紧型号实现竞争性性能,帮助放射科医生的工作负载.

关键词:

低级别的适应低级别的适应在ROCOv2数据集中,标题质量标题质量医疗图像标题标题医学图像标题具有参数效率的微调.放射学报告的生成视觉语言模型视觉语言模型

更多相关视频

Author Spotlight: Insights into Visual Cortex Research Through Wide-View fMRI Mapping

Author Spotlight: Insights into Visual Cortex Research Through Wide-View fMRI Mapping

Published on: December 8, 2023

相关实验视频

Last Updated: Jan 7, 2026

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness

Published on: December 6, 2024

Author Spotlight: Insights into Visual Cortex Research Through Wide-View fMRI Mapping

Author Spotlight: Insights into Visual Cortex Research Through Wide-View fMRI Mapping

Published on: December 8, 2023

科学领域:

医疗成像中的人工智能
计算机视觉计算机视觉
自然语言处理自然语言处理.

背景情况:

随着医疗成像数量的增加,需要自动化工具来支持放射科医生,并减少报告延迟.
视觉语言模型 (VLMs) 显示了通过自动化标题生成加速医疗报告起草的潜力.
用不同的参数尺度对VLM进行系统评估对于评估放射学中的临床效用至关重要.

研究的目的:

评估和比较十个多式视觉语言模型 (VLMs) 的性能,这些模型在大型医学成像数据集 (ROCOv2) 上进行了微调.
评估模型规模 (大VLM与小VLM) 和适应策略 (低级适应) 对医学成像中的VLM性能的影响.
建立定量基准来选择VLM用于医学成像解释的临床应用.

主要方法:

十个多式模式模型,包括大型VLM (LLaVA,IDEFICS-9B),小型VLM (MoonDream2,Qwen,SmolVLM) 和基线架构 (VisionGPT2,CNN-Transformer),在ROCOv2数据集上进行了微调 (116,635张图像,8种模式).
应用了低级调整 (LoRA),重点是通过最小的参数更新 (<1%的总参数) 优化性能.
模型使用相关性 (语义相似性) 和事实性 (概念级正确性) 的指标进行评估.

主要成果:

性能按模型尺度分层清晰:大VLM (0.273-0.317),小VLM (0.188-0.279) 和基线 (0.154-0.177).
LLaVA-Mistral-7B实现了最高的整体性能,显著超过了VisionGPT2基线.
月梦2 (小型VLM) 显示出竞争相关性得分,接近一些大型VLM的表现. 之前的模式标签显示了对小型VLM业绩的可变影响.

结论:

模型尺寸是医疗成像VLM性能的一个重要因素,大型VLM通常优异.
有针对性的适应策略,如优化的LoRA和特定的架构设计,使紧的小型VLM能够实现竞争性结果.
这些发现为VLM选择提供了重要的基准,突出了临床放射学工作流程中高效,较小的模型的潜力.