Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reducing self-supervised learning complexity improves weakly-supervised classification performance in computational pathology

Tim Lenz, Omar S. M. El Nahhas, Marta Ligero, Jakob Nikolas Kather

TL;DR

The paper addresses the barrier of heavy resource demands in self-supervised learning for computational pathology. It systematically reduces SSL complexity via data-volume reductions, encoder-stage modifications, and sampling strategy innovations within MoCo-v3 on a Swin Transformer backbone. Key contributions include showing that 50% SSL data suffices for downstream gene mutation tasks, demonstrating gains from multi-stage feature fusion, and introducing dynamic/negative sampling that outperform semantically relevant sampling, all while reducing training time by up to 90%. These findings enable effective foundation-model-style SSL for breast cancer histopathology on consumer hardware, broadening access and reducing costs for medical centers.

Abstract

Deep Learning models have been successfully utilized to extract clinically actionable insights from routinely available histology data. Generally, these models require annotations performed by clinicians, which are scarce and costly to generate. The emergence of self-supervised learning (SSL) methods remove this barrier, allowing for large-scale analyses on non-annotated data. However, recent SSL approaches apply increasingly expansive model architectures and larger datasets, causing the rapid escalation of data volumes, hardware prerequisites, and overall expenses, limiting access to these resources to few institutions. Therefore, we investigated the complexity of contrastive SSL in computational pathology in relation to classification performance with the utilization of consumer-grade hardware. Specifically, we analyzed the effects of adaptations in data volume, architecture, and algorithms on downstream classification tasks, emphasizing their impact on computational resources. We trained breast cancer foundation models on a large public patient cohort and validated them on various downstream classification tasks in a weakly supervised manner on two external public patient cohorts. Our experiments demonstrate that we can improve downstream classification performance whilst reducing SSL training duration by 90%. In summary, we propose a set of adaptations which enable the utilization of SSL in computational pathology in non-resource abundant environments.

Reducing self-supervised learning complexity improves weakly-supervised classification performance in computational pathology

TL;DR

The paper addresses the barrier of heavy resource demands in self-supervised learning for computational pathology. It systematically reduces SSL complexity via data-volume reductions, encoder-stage modifications, and sampling strategy innovations within MoCo-v3 on a Swin Transformer backbone. Key contributions include showing that 50% SSL data suffices for downstream gene mutation tasks, demonstrating gains from multi-stage feature fusion, and introducing dynamic/negative sampling that outperform semantically relevant sampling, all while reducing training time by up to 90%. These findings enable effective foundation-model-style SSL for breast cancer histopathology on consumer hardware, broadening access and reducing costs for medical centers.

Abstract

Deep Learning models have been successfully utilized to extract clinically actionable insights from routinely available histology data. Generally, these models require annotations performed by clinicians, which are scarce and costly to generate. The emergence of self-supervised learning (SSL) methods remove this barrier, allowing for large-scale analyses on non-annotated data. However, recent SSL approaches apply increasingly expansive model architectures and larger datasets, causing the rapid escalation of data volumes, hardware prerequisites, and overall expenses, limiting access to these resources to few institutions. Therefore, we investigated the complexity of contrastive SSL in computational pathology in relation to classification performance with the utilization of consumer-grade hardware. Specifically, we analyzed the effects of adaptations in data volume, architecture, and algorithms on downstream classification tasks, emphasizing their impact on computational resources. We trained breast cancer foundation models on a large public patient cohort and validated them on various downstream classification tasks in a weakly supervised manner on two external public patient cohorts. Our experiments demonstrate that we can improve downstream classification performance whilst reducing SSL training duration by 90%. In summary, we propose a set of adaptations which enable the utilization of SSL in computational pathology in non-resource abundant environments.
Paper Structure (17 sections, 6 equations, 1 figure, 3 tables)

This paper contains 17 sections, 6 equations, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Contrastive Self-Supervised Learning
  4. Semantically Relevant Sampling
  5. Positive Sampling.
  6. Negative Sampling.
  7. Dynamic Sampling.
  8. Swin Transformer Modifications
  9. Evaluation
  10. Experimental Results and Discussion
  11. Data.
  12. Our moco encoder improves over sota results on breast cancer gene classification tasks.
  13. A 50% reduction in SSL training data does not affect downstream gene mutation prediction performance.
  14. Incorporating feature representations from previous blocks improves weakly supervised learning performance.
  15. Negative sampling and dynamic sampling improve over semantically relevant contrastive learning.
  16. ...and 2 more sections

Figures (1)

  • Figure 1: Experiment Overview. (A) visualizes the contrastive SSL algorithm. (B) shows an overview of the experiments we performed in this study. (1) We train on subsets of the ssl data to analyze the impact of reduced data volumes during SSL on downstream classification performance (DCP). (2) We investigate the DCP of feature vectors retrieved from earlier stages of our SSL trained tiny Swin Transformer as well as combined feature representations of multiple layers. (3) We propose semantically relevant sampling methods to improve contrastive SSL in histopathology. Hence, for each feature representation of an input sample we rank all other representations of a batch based on the cosine similarity. Subsequently, we retrieve certain subsets of this ranking and adapt their weight in the contrastive loss function.