Overview of the Potentials of Multiple Instance Learning in Cancer Diagnosis: Applications, Challenges, and Future Directions

Abstract

The outcome of cancer patients mostly depends on the diagnosis process and the treatment strategies. Computer-aided diagnosis (CAD) methods have demonstrated the potential to handle accurate diagnostics using artificial intelligence techniques such as machine learning and deep learning. The nature of the data used in training the AI-based model determined the paradigm, often classified as supervised and unsupervised learning for scenarios with labeled and unlabeled data, respectively. Due to the cost of time and resources, most datasets are nowadays partially labeled and used for training. The weakly supervised learning approach enables the AI models to be trained with incompletely labeled, noisy, or imbalanced data. In recent years, multiple instance learning (MIL) has emerged as a promising weakly supervised learning approach in many fields, including cancer diagnosis. Unlike traditional supervised learning methods, MIL allows the classification of groups of instances, known as bags, where only the bag's label is available. This comprehensive review aims to provide an in-depth analysis of the applications of MIL in cancer diagnostic tasks, highlighting its advantages, challenges, and future directions. By examining these advantages, challenges, and future trends, the review aims to contribute to advancing MIL as a powerful tool in improving cancer diagnostic accuracy and patient outcomes.