Point-Beyond-Class（PBC）

胸部x线片 (CXR) 中准确的异常定位有助于各种胸部疾病的临床诊断。但是，病变级别的注释只能由经验丰富的放射科医生执行，并且繁琐且耗时，因此很难获得。这种情况导致难以为CXR开发完全监督的异常定位系统。在这方面，我们建议通过一种弱半监督策略来训练CXR异常定位框架，称为超越类点 (PBC)，该策略利用少量带有病变级边界框的完全注释的CXR和大量弱注释的样本逐点。这样的点注释设置可以以边际注释成本为异常定位提供弱实例级信息。特别是，我们PBC背后的核心思想是学习从点注释到边界框的可靠且准确的映射，以抵抗注释点的方差。为此，提出了一个正则化项，即多点一致性，该项驱动模型从同一异常内部的不同点注释生成一致的边界框。此外，还提出了一种称为对称一致性的自我监督，以深入利用弱注释数据中的有用信息进行异常定位。Accurate abnormality localization in chest X-rays (CXRs) aids clinical diagnosis of various thoracic diseases. However, lesion-level annotations can only be performed by experienced radiologists, and are tedious and time-consuming, making them difficult to obtain. This scenario makes it challenging to develop fully supervised abnormality localization systems for CXRs. In this regard, we propose training a CXR abnormality localization framework via a weakly semi-supervised strategy named Point Beyond Class (PBC), which leverages a small number of fully annotated CXRs with lesion-level bounding boxes and a large volume of weakly annotated samples at the point-level. This kind of point annotation setup can provide weak instance-level information for abnormality localization at marginal annotation costs. Specifically, the core idea behind our PBC is to learn a reliable and accurate mapping from point annotations to bounding boxes to resist the variance of annotated points. To this end, a regularization term named multi-point consistency is proposed, which drives the model to generate consistent bounding boxes from different point annotations within the same abnormality. In addition, a self-supervised method named symmetric consistency is proposed to deeply exploit useful information in weakly annotated data for abnormality localization.

搜集汇总

背景与挑战

背景概述

该数据集旨在通过一种名为PBC的弱半监督策略，解决胸部X光片异常定位中因全标注成本高昂而难以训练的问题。它利用少量带边界框的全标注样本和大量点标注样本来训练模型，其核心方法包括多点一致性和对称一致性正则化，以提升定位准确性并降低对标注的依赖。