使用卷积神经网络预测经腹超声图像中胃肠道间质瘤的恶性潜能：预测模型的焦点可视化。

Use of a Convolutional Neural Network to Predict the Malignant Potential of Gastrointestinal Stromal Tumors in Transabdominal Ultrasound Images: Visualization of the Focus of the Prediction Model.

机构信息

Department of Ultrasound, Fujian Medical University Union Hospital, Fuzhou, Fujian, China.

Department of General Surgery, Fujian Medical University Provincial Clinical Medical College, Fujian Provincial Hospital, Fuzhou, Fujian, China.

出版信息

Ultrasound Med Biol. 2023 Sep;49(9):1951-1959. doi: 10.1016/j.ultrasmedbio.2023.04.011. Epub 2023 Jun 7.

DOI:10.1016/j.ultrasmedbio.2023.04.011

PMID:37291007

Abstract

OBJECTIVE

We established a deep convolutional neural network (CNN) model based on ultrasound images (US-CNN) for predicting the malignant potential of gastrointestinal stromal tumors (GISTs).

METHODS

A total of 980 ultrasound images from 245 pathology-confirmed GIST patients after surgical operation were retrospectively collected and divided into a low (very-low-risk, low-risk) and a high (medium-risk, high-risk) malignant potential group. Eight pre-trained CNN models were used to extract the features. The CNN model with the highest accuracy in the test set was selected. The model's performance was evaluated by calculating accuracy, sensitivity, specificity, positive-predictive value (PPV), negative-predictive value (NPV) and the F1 score. Three radiologists with different experience levels also predicted the malignant potential of GISTs in the same test set. US-CNN and human assessments were compared. Subsequently, gradient-weighted class activation diagrams (Grad-CAMs) were used to visualize the model's final classification decisions.

RESULTS

Among the eight transfer learning-based CNNs, ResNet18 performed best. The accuracy, sensitivity, specificity, PPV, NPV and F1 score were 0.88, 0.86, 0.89, 0.82, 0.92 and 0.90, respectively, which were significantly better than those achieved by radiologists (resident doctor: 0.66, 0.55, 0.79, 0.74, 0.62 and 0.69; attending doctor: 0.68, 0.59, 0.78, 0.70, 0.69 and 0.73; professor: 0.69, 0.63, 0.72, 0.51, 0.80 and 0.76). Model interpretation with Grad-CAMs revealed that the activated areas mainly focused on cystic necrosis and margins.

CONCLUSION

The US-CNN model predicts GIST malignant potential well, which can assist in clinical treatment decision-making.

摘要

目的

我们建立了一个基于超声图像（US-CNN）的深度卷积神经网络（CNN）模型，用于预测胃肠道间质瘤（GIST）的恶性潜能。

方法

回顾性收集了 245 例经手术病理证实的 GIST 患者的 980 张超声图像，并将其分为低（极低风险、低风险）和高（中风险、高风险）恶性潜能组。使用 8 种预训练的 CNN 模型提取特征。选择测试集中准确率最高的 CNN 模型。通过计算准确率、敏感度、特异度、阳性预测值（PPV）、阴性预测值（NPV）和 F1 评分来评估模型的性能。3 名不同经验水平的放射科医生也在同一测试集中预测了 GIST 的恶性潜能。比较了 US-CNN 和人类评估。随后，使用梯度加权类激活图（Grad-CAMs）可视化模型的最终分类决策。

结果

在 8 种基于迁移学习的 CNN 中，ResNet18 表现最佳。其准确率、敏感度、特异度、PPV、NPV 和 F1 评分分别为 0.88、0.86、0.89、0.82、0.92 和 0.90，明显优于放射科医生（住院医师：0.66、0.55、0.79、0.74、0.62 和 0.69；主治医师：0.68、0.59、0.78、0.70、0.69 和 0.73；教授：0.69、0.63、0.72、0.51、0.80 和 0.76）。使用 Grad-CAMs 进行模型解释显示，激活区域主要集中在囊变坏死和边缘。