作者:Dinkar Juyal Siddhant Shingi Syed Ashar Javed Harshith Padigela Chintan Shah Anand Sampat Archit Khosla John Abel Amaro Taylor-Weiner
多实例学习(MIL)模型已在病理学中广泛用于预测生物标志物,并根据千兆像素大小对患者进行风险分层。医学成像中的机器学习问题通常涉及罕见病,这使得这些模型在标签不平衡设置中工作很重要。此外,当模型部署在现实世界中时,这些不平衡可能发生在分布外(OOD)数据集中。我们利用解耦特征和分类器学习的理念,可以改善标签不平衡数据集的决策基础。为此,我们研究了监督对比学习与多实例学习(SC-MIL)的整合。具体来说,我们提出了一个在标签不平衡的情况下的联合训练MIL框架,该框架从学习标签级表示逐渐过渡到最优分类器学习。我们对癌症病理学中两个研究充分的问题进行了不同失衡设置的实验:
Multiple Instance learning (MIL) models have been extensively used inpathology to predict biomarkers and risk-stratify patients from gigapixel-sizedimages. Machine learning problems in medical imaging often deal with rarediseases, making it important for these models to work in a label-imbalancedsetting. Furthermore, these imbalances can occur in out-of-distribution (OOD)datasets when the models are deployed in the real-world. We leverage the ideathat decoupling feature and classifier learning can lead to improved decisionboundaries for label imbalanced datasets. To this end, we investigate theintegration of supervised contrastive learning with multiple instance learning(SC-MIL). Specifically, we propose a joint-training MIL framework in thepresence of label imbalance that progressively transitions from learningbag-level representations to optimal classifier learning. We performexperiments with different imbalance settings for two well-studied problems incancer pathology: subtyping of non-small cell lung cancer and subtyping ofrenal cell carcinoma. SC-MIL provides large and consistent improvements overother techniques on both in-distribution (ID) and OOD held-out sets acrossmultiple imbalanced settings.
论文链接:http://arxiv.org/pdf/2303.13405v1
更多计算机论文:http://cspaper.cn/