Supervision and Source Domain Impact on Representation Learning: A Histopathology Case Study

Abstract

As many algorithms depend on a suitable representation of data, learning unique features is considered a crucial task. Although supervised techniques using deep neural networks have boosted the performance of representation learning, the need for a large sets of labeled data limits the application of such methods. As an example, high-quality delineations of regions of interest in the field of pathology is a tedious and time-consuming task due to the large image dimensions. In this work, we explored the performance of a deep neural network and triplet loss in the area of representation learning. We investigated the notion of similarity and dissimilarity in pathology whole-slide images and compared different setups from unsupervised and semi-supervised to supervised learning in our experiments. Additionally, different approaches were tested, applying few-shot learning on two publicly available pathology image datasets. We achieved high accuracy and generalization when the learned representations were applied to two different pathology datasets.

Batch-Incremental Triplet Sampling for Training Triplet Networks Using Bayesian Updating Theorem

“…The large CRC dataset includes nine classes of tissues, namely adipose, background, debris, lymphocytes, mucus, smooth muscle, normal colon mucosa (normal), cancer-associated stroma, and colorectal adenocarcinoma epithelium (tumor). Note that literature has shown the effectiveness of triplet variants networks for histopathology data, both with triplet-loss [26] and with NCA loss [27]; this shows the importance of validating our approaches on this domain.…”

Spectral, Probabilistic, and Deep Metric Learning: Tutorial and Survey

“…Various loss functions exist for supervised metric learning by neural networks. Supervised loss functions can teach the network to separate classes in the embedding space (Sikaroudi et al, 2020b). For this, we use a network whose last layer is for classification of data points.…”

“…where (y i ) l and f o (x i ) l denote the l-th element of y i and f o (x i ), respectively. Minimizing this loss separates classes for classification; this separation of classes also gives us discriminating embedding in the one-to-last layer (Sikaroudi et al, 2020b;Boudiaf et al, 2020).…”

Towards better understanding and better generalization of few-shot classification in histology images with contrastive learning

“…In the regard of histology image, Medela et al (2019) use a triplet loss (Schroff et al, 2015) to pre-train an encoder with a followed fine-tuned SVM classifier for few-shot domain adaptation. Sikaroudi et al (2020) and Teh & Taylor (2020) study learning with less data in histology images. Concurrent to our work, Shakeri et al (2021) simultaneously propose a benchmark for few-shot classification of histological images.…”