Page 53 - 南京医科大学自然版
P. 53

第46卷第4期          吴邦钰，王志霄，项景轩，等. 基于卷积神经网络的胃癌病理图像分类诊断与分级识别［J］.
                  2026年4月                      南京医科大学学报（自然科学版），2026，46（4）：520-532                       ·521 ·


                technology，and to evaluate the performance of these models. Methods：Classification and grading datasets for gastric cancer and non⁃
                cancerous tissues were collected from public online resources. Data augmentation was performed，and the dataset were divided into
                training，validation，and test sets. In the initial stage，17 convolutional neural network（CNN）architectures were constructed，and the
                initial training parameters were uniformly set to train these 17 models for the classification of gastric cancer and non⁃cancerous tissues.
                After training，the recognition accuracy on the test set and the training time were used as evaluation indicators to comprehensively
                assess the efficacy of different model architectures. Based on these indicators，the optimal architecture was selected for further
                optimization and training to construct the gastric cancer classification model. After the completion of the classification model，the
                gastric cancer grading model was built based on the foundation of the classification model. During the training of the gastric cancer
                grading model，17 grading networks were trained，and suitable base models were selected according to performance indicators. After
                the base model was determined，voting and stacking methods were applied for ensemble learning and compared with single models to
                explore the impact of ensemble learning on performance improvement and to construct the gastric cancer grading model. Results：In
                the training of the gastric cancer classification model，the Xception network was selected as the final classification model after
                comparison. After parameter adjustment and training，the final gastric cancer classification model achieved an accuracy of 98.13%，
                sensitivity of 98.11%，specificity of 98.11%，F1 score of 98.12%，and AUC of 0.998 on the test set. In the training of the gastric cancer
                grading model，the stacking method represented by random forest showed significant improvement compared to the voting method
                represented by hard voting. The ensemble model based on random forest was selected as the final grading model，with an accuracy of
                95.06%，sensitivity of 94.77%，specificity of 98.36%，and F1 score of 94.82%. The area under the receiver operating characteristic
               （ROC ⁃ AUC）curve values were 0.999 for benign，0.981 for poorly differentiated tubular adenocarcinoma，0.990 for moderately
                differentiated tubular adenocarcinoma，and 0.995 for well ⁃ differentiated tubular adenocarcinoma. Conclusion：Both models
                demonstrated excellent recognition performance，proving the feasibility of using CNN to achieve high ⁃ precision classification and
                grading of gastric tumor pathological images. The transfer⁃learning and ensemble⁃learning framework was successfully applied to the
                grading of gastric tumor images and holds promise for integration into hospital intelligent diagnostic assistance systems.
               ［Key words］ gastric cancer；precision medicine；convolutional neural network；ensemble learning；artificial intelligence
                                                                              ［J Nanjing Med Univ，2026，46（04）：520⁃532］






                    胃癌是全球的重大公共卫生挑战之一。根据                               近年来，AI技术在医学图像分析中取得突破性
                国 际 癌 症 研 究 机 构（International Agency for Re⁃      进展  ［4- 5］ 。卷积神经网络（convolutional neural net⁃
                                                                                                          ［6］
                search on Cancer）发布的 GLOBOCAN 2022 数据，全           work，CNN）通过端到端特征学习，在乳腺癌 、肺
                                                                    ［7］
                球胃癌年新增病例约 109 万例，死亡 76 万例，发病                      癌 等病理图像分类任务中展现出超越传统方法的
                                                                                                     ［8］
                率与死亡率分别位居癌症谱第5位和第4位。东亚                            性能。轻量化架构如 EfficientNet⁃V2 与 Mobile⁃
                                                                       ［9］
                地区（中国、日本、韩国）的年龄标准化发病率（age⁃                        NetV3 通过复合缩放策略，在保持高精度的同时显
                standardized rate of incidence，ASR）高达 24.3/10 万，  著降低计算复杂度，为边缘设备部署奠定基础。然
                显著高于全球平均水平（11.1/10万），这一差异与幽                       而，胃癌病理图像的异质性（如分化程度、染色差
                门螺杆菌高感染率（>50%）、高盐饮食及遗传多态性                         异）对模型泛化能力提出更高要求，亟需系统性探
               （如CDH1基因突变）密切相关 。中国国家癌症中                           索多模型集成与迁移学习的协同优化策略 。
                                           ［1］
                                                                                                       ［10］
                心 2023 年统计显示，胃癌新发病例 47.9 万例，死亡                        本研究通过系统性评估 17 种 CNN 模型在胃癌
                37.4 万例，分别占恶性肿瘤发病和死亡的第 3 位与                       病理图像分类诊断与分级任务中的性能，筛选最优
                第2位，且约80%患者确诊时已进展至中晚期，5年生                         架构并构建集成学习框架。在胃癌诊断中引入多
                            ［2］
                存率仅为28% 。由于早期胃癌缺乏特异性症状，现                          模型横向对比，量化不同网络在特征提取能力、训
                行诊断体系高度依赖内镜活检与病理学评估，但传统                           练效率及泛化性能方面的差异。并提出一种基于
                病理诊断受限于医师经验差异与视觉疲劳，误诊率可                           Stacking 策略的异构模型集成方法，通过多模型集
                达 8.94% 。因此，开发高精度、标准化的人工智能                        成进行互补特征融合，提高模型准确率与泛化性。
                       ［3］
               （artificial intelligence，AI）辅助诊断工具对提升胃癌            以期通过迁移⁃集成联合框架，促进 AI 模型在低资
                早筛效率和减少误诊率具有重要临床价值。                               源医疗场景下的临床快速部署，为数字化病理诊断
   48   49   50   51   52   53   54   55   56   57   58