Towards Multiple Instance Learning in Computational Pathology: Methodologies, Applications, Challenges, Recent Trends and Future Directions

Multiple Instance Learning (MIL) is a powerful variant in the machine learning paradigm, which typically deals with incomplete or weakly labeled data. In traditional supervised learning, for every training sample, there is a ground truth label; however, in MIL, labels are associated with bags of instances, where each bag consists of many instances. MIL is beneficial where data labeling is costly in terms of time and resources. In computational pathology, it helps deal with complex and large-sized WSIs where labels are available for whole-slide images. To annotate every pixel of WSI is challenging, error-prone, and costly. This paper investigates a multiple instance-based approach for lung cancer classification. Moreover, it explores the potential of multiple instance-based approaches for cancer diagnosis in WSIs. Furthermore, it investigates different applications in computational pathology, algorithms, challenges, recent developments, and future prospects.