{"title":"基于 ResNet50 的有丝分裂细胞分类器的新型多级迭代训练策略","authors":"Yuqi Chen, Juan Liu, Peng Jiang, Yu Jin","doi":"10.1016/j.compbiolchem.2024.108092","DOIUrl":null,"url":null,"abstract":"<div><p>The number of mitotic cells is an important indicator of grading invasive breast cancer. It is very challenging for pathologists to identify and count mitotic cells in pathological sections with naked eyes under the microscope. Therefore, many computational models for the automatic identification of mitotic cells based on machine learning, especially deep learning, have been proposed. However, converging to the local optimal solution is one of the main problems in model training. In this paper, we proposed a novel multilevel iterative training strategy to address the problem. To evaluate the proposed training strategy, we constructed the mitotic cell classification model with ResNet50 and trained the model with different training strategies. The results showed that the models trained with the proposed training strategy performed better than those trained with the conventional strategy in the independent test set, illustrating the effectiveness of the new training strategy. Furthermore, after training with our proposed strategy, the ResNet50 model with Adam optimizer has achieved 89.26% F1 score on the public MITOSI14 dataset, which is higher than that of the state-of-the-art methods reported in the literature.</p></div>","PeriodicalId":10616,"journal":{"name":"Computational Biology and Chemistry","volume":null,"pages":null},"PeriodicalIF":2.6000,"publicationDate":"2024-05-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"A novel multilevel iterative training strategy for the ResNet50 based mitotic cell classifier\",\"authors\":\"Yuqi Chen, Juan Liu, Peng Jiang, Yu Jin\",\"doi\":\"10.1016/j.compbiolchem.2024.108092\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><p>The number of mitotic cells is an important indicator of grading invasive breast cancer. It is very challenging for pathologists to identify and count mitotic cells in pathological sections with naked eyes under the microscope. Therefore, many computational models for the automatic identification of mitotic cells based on machine learning, especially deep learning, have been proposed. However, converging to the local optimal solution is one of the main problems in model training. In this paper, we proposed a novel multilevel iterative training strategy to address the problem. To evaluate the proposed training strategy, we constructed the mitotic cell classification model with ResNet50 and trained the model with different training strategies. The results showed that the models trained with the proposed training strategy performed better than those trained with the conventional strategy in the independent test set, illustrating the effectiveness of the new training strategy. Furthermore, after training with our proposed strategy, the ResNet50 model with Adam optimizer has achieved 89.26% F1 score on the public MITOSI14 dataset, which is higher than that of the state-of-the-art methods reported in the literature.</p></div>\",\"PeriodicalId\":10616,\"journal\":{\"name\":\"Computational Biology and Chemistry\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":2.6000,\"publicationDate\":\"2024-05-10\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Computational Biology and Chemistry\",\"FirstCategoryId\":\"99\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S147692712400080X\",\"RegionNum\":4,\"RegionCategory\":\"生物学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"BIOLOGY\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Computational Biology and Chemistry","FirstCategoryId":"99","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S147692712400080X","RegionNum":4,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"BIOLOGY","Score":null,"Total":0}
引用次数: 0
摘要
有丝分裂细胞的数量是浸润性乳腺癌分级的一个重要指标。对于病理学家来说,在显微镜下用肉眼识别和计数病理切片中的有丝分裂细胞非常具有挑战性。因此,人们提出了许多基于机器学习,尤其是深度学习的有丝分裂细胞自动识别计算模型。然而,收敛到局部最优解是模型训练的主要问题之一。本文提出了一种新颖的多级迭代训练策略来解决这一问题。为了评估所提出的训练策略,我们用 ResNet50 构建了有丝分裂细胞分类模型,并用不同的训练策略对模型进行了训练。结果表明,在独立测试集上,用提出的训练策略训练的模型比用传统策略训练的模型表现更好,说明了新训练策略的有效性。此外,使用我们提出的策略训练后,带有 Adam 优化器的 ResNet50 模型在公开的 MITOSI14 数据集上取得了 89.26% 的 F1 分数,高于文献中报道的最先进方法。
A novel multilevel iterative training strategy for the ResNet50 based mitotic cell classifier
The number of mitotic cells is an important indicator of grading invasive breast cancer. It is very challenging for pathologists to identify and count mitotic cells in pathological sections with naked eyes under the microscope. Therefore, many computational models for the automatic identification of mitotic cells based on machine learning, especially deep learning, have been proposed. However, converging to the local optimal solution is one of the main problems in model training. In this paper, we proposed a novel multilevel iterative training strategy to address the problem. To evaluate the proposed training strategy, we constructed the mitotic cell classification model with ResNet50 and trained the model with different training strategies. The results showed that the models trained with the proposed training strategy performed better than those trained with the conventional strategy in the independent test set, illustrating the effectiveness of the new training strategy. Furthermore, after training with our proposed strategy, the ResNet50 model with Adam optimizer has achieved 89.26% F1 score on the public MITOSI14 dataset, which is higher than that of the state-of-the-art methods reported in the literature.
期刊介绍:
Computational Biology and Chemistry publishes original research papers and review articles in all areas of computational life sciences. High quality research contributions with a major computational component in the areas of nucleic acid and protein sequence research, molecular evolution, molecular genetics (functional genomics and proteomics), theory and practice of either biology-specific or chemical-biology-specific modeling, and structural biology of nucleic acids and proteins are particularly welcome. Exceptionally high quality research work in bioinformatics, systems biology, ecology, computational pharmacology, metabolism, biomedical engineering, epidemiology, and statistical genetics will also be considered.
Given their inherent uncertainty, protein modeling and molecular docking studies should be thoroughly validated. In the absence of experimental results for validation, the use of molecular dynamics simulations along with detailed free energy calculations, for example, should be used as complementary techniques to support the major conclusions. Submissions of premature modeling exercises without additional biological insights will not be considered.
Review articles will generally be commissioned by the editors and should not be submitted to the journal without explicit invitation. However prospective authors are welcome to send a brief (one to three pages) synopsis, which will be evaluated by the editors.