研究目的
To automate the classification process of mouse tissues into three classes: healthy, inflammation, and cancer tissues using machine learning approaches.
研究成果
The study presents three classification approaches to classify three states of health, inflammation, and cancer on mice colon’s wall, achieving high performance with fully automated machine learning-based methods. The best classification performance was achieved with a custom deep learning model, demonstrating the potential of machine learning in preclinical studies despite the small size of the database.
研究不足
The amount of data in preclinical studies is limited compared to the usual amount of data available in medical applications of machine learning. The shift in data distributions between natural images and medical images is too large for some pre-trained models.
1：Experimental Design and Method Selection:
The study employs fully automated machine learning-based methods, including deep learning, transfer learning, and shallow learning with SVM, to classify the health state of mice colon tissues.
2：Sample Selection and Data Sources:
The database consists of 66788 images from 38 mice, annotated as healthy, cancer, or inflammation tissue images.
3：List of Experimental Equipment and Materials:
Confocal endomicroscopy imaging protocol, mini multi-purpose rigid telescope, Proflex MiniZ microprobe, and a solution of Fluorescein Isothiocyanate FITC- Dextran 5% as a contrast agent.
4：Experimental Procedures and Operational Workflow:
Images were acquired using a confocal endomicroscope combined with a microprobe inserted through the operating sheath of an endoscope. The depth assessed was approximately 58 μm for a lateral resolution of
5：5 μm and a frame rate of 12 fps. Data Analysis Methods:
The study uses LBP features for handcrafted feature representation and deep convolutional neural networks for representation learning, with performance evaluated through cross-subject and cross-sample training strategies.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
