Analysis of Large Language Models for Magnetic Resonance Imaging Brain Tumor Segmentation Reporting

作者
Vennila M,Thiruvenkadam Kalaiselvi
标识
DOI:10.7759/s44389-026-00225-5
摘要

This paper builds an end-to-end model that automates the process of analyzing and reporting brain tumors based on multiparametric MRI scans, which is a step toward overcoming the bottleneck of translating specific imaging health records into specific clinical reports in neuroradiology. The pipeline takes post-treatment glioma cases from the Multimodal Brain Tumor Image Segmentation (BraTS) 2024 dataset, which has been divided into 80% for training and 20% for validation, and Swin UNETR is used for 3D segmentation of tumor subregions to improve the identification of the tumor, tumor core, and whole tumor, followed by the extraction of quantitative features of subregion volumes, geometric centroids, and radiomics descriptors. These structured queries prompt three large language models: GPT-5 for higher-level diagnostic reasoning, Claude 4 for safety-conscious, clinically aligned phrasing, and Llama 4 for concise, high-throughput triage reporting. The methodology entails using high-precision segmentation and valid multi-metric assessment, which consists of Dice similarity for visual consistency and BERTScore for semantic consistency, GREEN score for clinical safety, and RadGraph F1 for factual entity-relation overlap. The comparative analysis compares the models on the basis of diagnostic reasoning, liability-conscious documentation, and rapid triage use cases and indicates the task-specific advantages that can be used to direct clinical deployment. The framework bridges the image-to-text gap that has existed in neuroradiology, with multimodal integration providing production-scale artificial intelligence suitable for picture archiving and communication systems and demonstrating that model selection should be based on workflow priorities - depth of analysis versus generation speed. This method creates a clinically translational standard for neuro-oncology reporting, which can provide real-time support for treatment planning with quantitative confidence and informative results.

科研通智能强力驱动
Strongly Powered by AbleSci AI
科研通是完全免费的文献互助平台,具备全网最快的应助速度,最高的求助完成率。 对每一个文献求助,科研通都将尽心尽力,给求助人一个满意的交代。
实时播报
看文献完成签到,获得积分10
刚刚
chenying完成签到 ,获得积分0
4秒前
shuang完成签到,获得积分10
5秒前
5秒前
我就是KKKK完成签到,获得积分10
6秒前
gsxgg完成签到,获得积分10
8秒前
11秒前
快乐的小木虫完成签到,获得积分10
11秒前
csy完成签到,获得积分10
12秒前
fangyuan完成签到,获得积分10
13秒前
郑琦敏钰完成签到 ,获得积分10
14秒前
davyean完成签到,获得积分10
14秒前
aaa完成签到,获得积分10
14秒前
淡淡依霜完成签到 ,获得积分10
16秒前
辣条我有呀完成签到 ,获得积分10
16秒前
jiangsi完成签到 ,获得积分10
17秒前
酷炫的大碗完成签到,获得积分10
17秒前
Daisy完成签到 ,获得积分10
18秒前
执着的天使完成签到 ,获得积分10
19秒前
生动的乐曲完成签到,获得积分10
19秒前
ChengYonghui完成签到,获得积分10
20秒前
火星上蜜粉完成签到,获得积分10
21秒前
树妖三三完成签到,获得积分10
22秒前
22秒前
crab完成签到,获得积分10
23秒前
厉奕欣完成签到 ,获得积分10
24秒前
QwQ完成签到 ,获得积分10
27秒前
SSY发布了新的文献求助10
27秒前
Wei发布了新的文献求助20
30秒前
Orange的应助被酷酷的怀莲采纳,获得10
34秒前
36秒前
xfy的应助被科研通管家采纳,获得10
36秒前
淡然雁梅完成签到 ,获得积分10
37秒前
ri_290完成签到,获得积分10
38秒前
38秒前
xzcx完成签到 ,获得积分10
38秒前
方方完成签到,获得积分10
39秒前
心如止水完成签到,获得积分10
41秒前
SCINEXUS完成签到,获得积分0
41秒前
怡然绯完成签到 ,获得积分10
42秒前
高分求助中
(应助此贴封号)【重要!!请各用户(尤其是新用户)详细阅读】【科研通的精品贴汇总】 10000
拉伸高分子熔体中缠结行为的模拟和小角中子散射研究 1700
The World Atlas of Wine 9th edition 1000
奈特基础组织学:配套组织病理学(第4版)(Netter's Essential Histology With Correlated Histopathology) 800
John McDowells Bildungstheorie 800
八旗制度の研究 600
Hidden Empire of Finance: How Wall Street Profits from Our Cities and Fuels Global Inequality 500
热门求助领域 (近24小时)
化学 材料科学 医学 生物 纳米技术 计算机科学 工程类 有机化学 化学工程 生物化学 物理 复合材料 内科学 催化作用 光电子学 人工智能 细胞生物学 无机化学 物理化学 数学
热门帖子
关注 科研通微信公众号,转发送积分 8166556
求助须知:如何正确求助?哪些是违规求助? 9559576
关于积分的说明 21750809
捐赠科研通 7757392
什么是DOI,文献DOI怎么找? 3429167
关于科研通互助平台的介绍 2553530
邀请新用户注册赠送积分活动 2473923