Label Cleaning Multiple Instance Learning: Refining Coarse Annotations on Single Whole-slide Imaging

Annotating cancerous regions in whole-slide images (WSIs) plays a critical role in both clinical diagnosis, biomedical research, and machine learning algorithms development. However, exhaustive and accurate annotations are costly to obtain. Coarse annotations, such as rough boundaries for tumor, are much easier to obtain, which alleviate pathologists’ workload and save time for more urgent tasks. Although there are some studies on obtaining machine learning models from these inaccurate annotations, few of them tackle the refinement problem, where the mislabeled samples need to be explicitly identified and relabeled. In this thesis, we propose a method called Label Cleaning Multiple Instance Learning (LC-MIL) to refine annotations on a single WSI. Patches cropped from a WSI with noisy labels are packed into bags and an attention mechanism is leveraged to discriminate mislabeled instances and to mitigate their impact on bag-level prediction. Our experiments on a heterogeneous WSI set with lymph node metastasis, liver cancer, and colorectal cancer samples show that LC-MIL significantly refines the coarse annotations, even while learning from a single slide. To the best of my knowledge, this is the first time the MIL framework is used to handle label noise. The presented methodology could become especially useful since neither a large dataset nor tedious or expensive annotations are needed to implement it