Most recent advances in automated approaches towards WSI classification employ a two-stage pipeline, i.e. patch-level classification followed by WSI-level aggregation. The training of patch classifiers requires exact patch-level labels. This process either relies on exact labeled patches, which are time-consuming to acquire, or is based on the assumption that most patches of a WSI are discriminative for diagnosis. However, discriminative tumor regions often only occupy a small portion of WSIs. One existing method used a pre-trained CNN to extract patch-level features. The extracted features are further averaged to get a WSI representation. Directly pooling on all patches in each WSI undermines the discriminative power of rare tumor patches in classification.

In this thesis we p...[ Read more ]

Most recent advances in automated approaches towards WSI classification employ a two-stage pipeline, i.e. patch-level classification followed by WSI-level aggregation. The training of patch classifiers requires exact patch-level labels. This process either relies on exact labeled patches, which are time-consuming to acquire, or is based on the assumption that most patches of a WSI are discriminative for diagnosis. However, discriminative tumor regions often only occupy a small portion of WSIs. One existing method used a pre-trained CNN to extract patch-level features. The extracted features are further averaged to get a WSI representation. Directly pooling on all patches in each WSI undermines the discriminative power of rare tumor patches in classification.

In this thesis we propose a data-driven feature aggregation approach which could automatically identify discriminative features. Specifically, we first cluster all patches of each WSI into different clusters. Average pooling is applied among patches in each cluster. It helps remove redundant features without losing discriminative information for diagnosis. The clustering centroids are collected as instances in the WSI bag. Moreover, labels of patches cropped from normal WSIs are also normal. We learn a GMM on the normal patches. Features with low GMM scores are more likely to be tumor ones The learned GMM is used to select discriminative patches. Also the clustering centroids in each WSI are sorted based on their GMM scores. We adopt two methods to aggregate the cluster centroids: one is sequence-invariant MIL pooling methods, the other encodes the sequence information by IndRNN.

To demonstrate the effectiveness of our method, we apply it to predicting breast cancer metastases in lymph nodes, where only small regions in tumor WSIs have discriminative tumor patterns. To the best of our knowledge we are the first to analyze the data without supervised patch-level labels and achieved satisfactory results.