Core Concepts
Contrastive-based deep embeddings exhibit superior resilience to label noise compared to non-contrastive embeddings and image-based methods in histopathology image classification.
Abstract
The paper presents a comprehensive evaluation of the robustness of deep embeddings extracted from various pretrained histopathology foundation models under different label noise scenarios. The key findings are:
Classifiers trained on contrastive-based deep embeddings demonstrate improved robustness to label noise compared to those trained on the original images using state-of-the-art noise-resilient methods.
Contrastive-based embeddings exhibit superior noise tolerance compared to non-contrastive embeddings, even when the backbones are trained on unrelated domains like ImageNet.
The observed performance differences are not due to the quality of the learned representations, but rather to the noise-resilient property leveraged by the linear classifier when trained with contrastive embeddings.
While contrastive learning effectively mitigates the label noise challenge, it does not completely eliminate it, especially for relatively small datasets and under asymmetric noise scenarios. Further research is needed to develop methods that can fully overcome the label noise issue.
Stats
"Recent advancements in deep learning have proven highly effective in medical image classification, notably within histopathology."
"To be effective, training such deep neural networks (DNNs) requires large image datasets with reliable labels. However, in the context of medical imaging and histopathology in particular, clean data are rare and expensive, requiring expert labeling campaigns."
"These inaccuracies, stemming from inter-observer variability, imperfect segmentation of tissue regions, inherent ambiguity in the biological features, and omission errors, impede the development of reliable deep learning models."
"It has been proven that DNNs can easily overfit noisy labels (Li et al., 2018; Zhang et al., 2016), leading to severe degradations in model performance and thus potentially misleading clinical decisions."
Quotes
"We demonstrate that classifiers trained on contrastive deep embeddings exhibit improved robustness to label noise compared to those trained on the original images using state-of-the-art methods."
"Across nearly all the datasets and noise rates scenarios, these methods consistently match or surpass performances of image-based approaches."
"Particularly noteworthy is the observation that, for noise rates η > 0, classifiers trained with contrastive embeddings exhibit superior performance compared to their non-contrastive counterparts."