PO.BCS02.04 · 生物信息与计算
利用基础模型从内镜图像预测浅表性食管癌的浸润深度
Depth prediction in superficial esophageal cancer using a foundation model from endoscopic images
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：对于无淋巴结转移（LNM）风险的浅表性食管癌（SEC），首选的治疗方式是前期内镜下黏膜剥离术（ESD）而非手术。预测LNM风险的代表性因素是肿瘤浸润深度，但术前评估阶段预测准确性不足，常导致不必要的手术或ESD后需要挽救性手术。因此，我们提出利用应用于术前内镜图像的基础模型（FM）来预测SEC肿瘤浸润深度。我们使用在约500,000张EGD图像上预训练的基础模型（GastroFM），提出利用其从预训练中获得的知识，从内镜图像提供可靠的、组织病理学之前的浸润深度预测。
方法：我们改进了一个基础模型，该模型基于视觉Transformer架构，并在约500,000张上消化道内镜图像上预训练，用于SEC浸润深度预测。我们提出利用从该预训练中获得的知识，从内镜图像提供可靠的、组织病理学之前的浸润深度预测。该模型在一个包含839例SEC ESD病例（2007年4月-2023年1月）的回顾性队列上进行训练和验证，采用8:1:1的训练/验证/测试划分。对于每个病例，食管ESD专家从术前内镜中选取显示肿瘤的图像（中位数10张，范围4-36张）。我们应用基于注意力的多示例学习（AbMIL）将这些可变长度的图像序列聚合为单一的病例级预测评分，用于二分类任务：黏膜癌 vs 黏膜下癌。
结果：整体队列分布为529例（63.1%）黏膜癌和310例（36.9%）黏膜下癌。在测试集上，FM取得了0.821的AUC和0.810的总体准确率。二分类的关键指标为：敏感性0.645、特异性0.906、F1评分0.714、阳性预测值0.800、阴性预测值0.814。
结论：我们开发了一个对SEC浸润深度预测具有判别力的FM。该模型有望成为术前评估中一个有价值的辅助工具，协助内镜医师和外科医师为SEC确定最佳治疗方案。
查看英文原文 English abstract
Background: For superficial esophageal cancer (SEC) without a risk of lymph node metastasis (LNM), upfront endoscopic submucosal dissection (ESD) is the preferred treatment over surgery. The representative factor for predicting LNM risk is tumor depth, but the insufficient prediction accuracy during pre-procedure evaluation frequently results in unnecessary surgery or the need for salvage surgery after ESD. Accordingly, we propose to predict SEC tumor depth using a foundation model (FM) applied to pre-procedural endoscopic images. Using a foundation model (GastroFM) that was pretrained on approximately 500,000 EGD images, we propose utilizing its knowledge gained from pre-training to provide a reliable, pre-histopathology prediction of invasion depth from endoscopic images.
Methods: We adapted a FM, which is based on a vision transformer architecture and was pretrained on approximately 500,000 upper endoscopic images, for SEC depth prediction. We propose utilizing the knowledge gained from this pre-training to provide a reliable, pre-histopathology prediction of invasion depth from endoscopic images. The model was trained and validated on a retrospective cohort of 839 ESD cases for SEC (April 2007-January 2023) using an 8:1:1 train/validation/test split. For each case, esophageal ESD expert selected tumor-displaying images (median 10, range 4-36) from pre-procedure endoscopy. We applied Attention-based Multiple Instance Learning (AbMIL) to aggregate these variable-length image sequences into a single case-level prediction score for the binary task: mucosal cancer vs submucosal cancer.
Results: The overall cohort distribution was 529 cases (63.1%) with mucosal cancer and 310 cases (36.9%) with submucosal cancer. On the test set, a FM achieved an AUC of 0.821 and an overall accuracy of 0.810. Key metrics for the binary classification were: sensitivity 0.645, specificity 0.906, F1 score 0.714, positive predictive value 0.800, and negative predictive value 0.814.
Conclusion: We developed a FM with a discriminative power for depth prediction in SEC. This model is expected to be a valuable supplementary tool in the pre-procedure workup, assisting endoscopists and surgeons in determining the optimal treatment plan for SEC.
利益披露 Disclosure
S. Kim, None..
S. Kim, None..
H. Yoo, None..
H. Lee, None..
Y. Min, None.