可解释性
杠杆(统计)
计算机科学
人工智能
深度学习
排名(信息检索)
机器学习
上下文图像分类
模式识别(心理学)
数据挖掘
图像(数学)
标识
DOI:10.1109/isbi52829.2022.9761678
摘要
Whole slide scanning is a powerful tool in clinical diagnosis and pathological research. However, it’s time-consuming to acquire localized annotations in whole slide images (WSIs). Recently, deep multiple instance learning (MIL) approaches were proposed to classify WSIs with only global annotations. Two main challenges, interpretability and utilizing multiple-scale information, remain to be solved in these approaches. In this study, we proposed a deep hierarchical multiple in-stance learning model to tackle these challenges. We introduced max-max ranking loss to better leverage the standard MIL assumption for better interpretability. A hierarchical architecture was designed to reduce computational costs and to utilize multiple-scale information. Our model was evaluated in a large WSI dataset CAMELYON16 with accuracy and AUC as metrics. Experimental results showed that our model achieved the best performance.
科研通智能强力驱动
Strongly Powered by AbleSci AI