English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

The synergy of long-range dependencies from transformers and localrepresentations of image content from convolutional neural networks (CNNs) hasled to advanced architectures and increased performance for various medicalimage analysis tasks due to their complementary benefits. However, comparedwith CNNs, transformers require considerably more training data, due to alarger number of parameters and an absence of inductive bias. The need forincreasingly large datasets continues to be problematic, particularly in thecontext of medical imaging, where both annotation efforts and data protectionresult in limited data availability. In this work, inspired by the humandecision-making process of correlating new evidence with previously memorizedexperience, we propose a Memorizing Vision Transformer (MoViT) to alleviate theneed for large-scale datasets to successfully train and deploytransformer-based architectures. MoViT leverages an external memory structureto cache history attention snapshots during the training stage. To preventoverfitting, we incorporate an innovative memory update scheme, attentiontemporal moving average, to update the stored external memories with thehistorical moving average. For inference speedup, we design a prototypicalattention learning method to distill the external memory into smallerrepresentative subsets. We evaluate our method on a public histology imagedataset and an in-house MRI dataset, demonstrating that MoViT applied to variedmedical image analysis tasks, can outperform vanilla transformer models acrossvaried data regimes, especially in cases where only a small amount of annotateddata is available. More importantly, MoViT can reach a competitive performanceof ViT with only 3.0% of the training data.

MoViT: Memorizing Vision Transformers for Medical Image Analysis