Prototypical Few-Shot Learning for Histopathology Classification: Leveraging Foundation Models With Adapter Architectures
Histopathology is a critical tool for disease diagnosis and identifying cancer via microscopic tissue analysis. Traditional deep learning methods for histopathology often require extensive labeled data, which can be scarce and expensive. This study introduces a framework for the few-shot adaptation...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11005525/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Histopathology is a critical tool for disease diagnosis and identifying cancer via microscopic tissue analysis. Traditional deep learning methods for histopathology often require extensive labeled data, which can be scarce and expensive. This study introduces a framework for the few-shot adaptation of self-supervised histopathology pretrained foundation models using multilayer perception adapters and convolutional adapters. An adapter comprising two linear or convolutional layers with nonlinear activation and residual connections transforms embeddings from foundation models for histopathology classification tasks. This study employs prototypical networks, SimpleShot, and bias-diminishing cosine similarity-based prototypical networks as few-shot learning algorithms. Comprehensive experiments are conducted across benchmark histopathology datasets: NCT, LC25000, Kather, and Camelyon17 Wilds. The results demonstrate that both adapter architectures consistently outperform the linear probe method, whereas multilayer perception adapters have an overall higher accuracy, especially when fine-tuned with five or more samples. The iBOT model with multilayer perception adapters fine-tuned using the bias-diminishing cosine similarity-based prototypical network algorithm achieved remarkable accuracy, reaching 95.04% on Camelyon17 Wilds and 96.55% on the NCT dataset with 20 images per class while using less than 0.002% of the dataset. These findings underscore the effectiveness of the proposed approach in addressing challenges posed by low-data regimes in the computer-aided histopathology domain and the potential for optimizing foundation models with minimal labeled data using prototypical few-shot algorithms. |
|---|---|
| ISSN: | 2169-3536 |