Deep Learning in Multiple Instance Learning: Methods, Applications, and Research Trends

Multiple instance learning (MIL) is a flexible learning framework where data is organized into sets of instances, called bags, with a single label assigned to the entire bag instead of individual instances. This formulation is especially suitable for addressing complex problems by allowing greater flexibility in their representation using weakly labeled data. As in other learning frameworks, Deep Learning (DL) approaches have been applied to solve MIL problems showing very interesting results. This systematic review aims to provide a comprehensive review of all recent research efforts on DL based on MIL. Thus, 173 papers published from 2016 are analyzed. A unified taxonomy is proposed that relates DL architectures with MIL strategies and learning frameworks, providing an integrated view on their interactions. In addition, the main types of architectures, learning frameworks, assumptions, applications, datasets, and tools are discussed in detail. Relevant findings indicate that Convolutional Neural Networks remain the most commonly used architectures, with a growing trend toward hybrid models that incorporate attention mechanisms, transformers, or recurrent networks. While early MIL approaches primarily relied on instance-based strategies under the standard assumption, more recent studies reveal a clear shift toward the extended assumption, particularly within hybrid architectures. Bag-based strategies remain less prevalent and are mainly explored through hybrid models. Most of the reviewed studies focus on classification tasks, where the diversity of MIL assumptions and architectural choices is more pronounced, while regression tasks have been addressed less frequently. Regarding the applications, medical diagnosis is the most widely used application, with the most available resources. Also, it was found that Python is the most frequently used tool in the reviewed papers. Finally, we point out the open issues and future research directions. Some key challenges such as the adaptation of generative architectures in unsupervised learning and the improvement of robustness in the models are still open to be addressed.