Research map / Label-efficient and robust learning / Domain adaptation and generalization
Memory-Efficient Prompt Tuning for Incremental Histopathology Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a memory-efficient prompt tuning framework for domain-incremental histopathology classification, using per-domain domain-specific prompts stored in a bank and a domain-invariant prompt refined via a style-augmented graph attention network.
Problem
Domain incremental learning for histopathology classification must avoid catastrophic forgetting and generalize to unseen domains under privacy constraints (no past data, no domain identity) while keeping memory cost low for gigapixel images.
Approach
Freeze a ViT-B/16 backbone and attach two lightweight prompts per time step: a domain-specific prompt stored and isolated in a prompt bank, and a domain-invariant prompt evolved across time by a graph attention network over prompt nodes, trained on Fourier amplitude-style-augmented images.
Key results
- On Camelyon17 the method reaches IL 93.67, BWT -1.62 and FTU 82.17, the best in 8 of 9 metrics versus S-Prompt (IL 91.73, FTU 80.64).
- On epithelium-stroma classification it reaches average final Acc 90.52 and IL 92.17, versus DualPrompt at 88.45 and 91.14.
- Memory overhead is about 0.57 MB additional storage with model size efficiency 0.99, comparable to other prompt-based methods, while replay and isolation methods such as Orc-MML need about 460 MB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply decoupled prompt tuning with style-augmented prompt refining | ViT-B/16 for incremental histopathology classification · Camelyon17 Domain 5 (unseen) | accuracy: 5.55% | Helped |
| Apply decoupled prompt tuning with style-augmented prompt refining | ViT-B/16 for incremental histopathology classification · Camelyon17 | FTU: 1.53% | Helped |
| Apply decoupled prompt tuning with style-augmented prompt refining | ViT-B/16 for epithelium-stroma classification · Epithelium-stroma Domain 4 (unseen) | accuracy: 2.42% | Helped |
| Apply decoupled prompt tuning with style-augmented prompt refining | ViT-B/16 for epithelium-stroma classification · Epithelium-stroma | FTU: 1.77% | Helped |
| Add a domain-specific prompt to the domain-invariant prompt (ablation) | Incremental prompt tuning · Camelyon17 | A-Acc: 85.94 -> 86.24 | Helped |
| Add a domain-invariant prompt to the domain-specific prompt (ablation) | Incremental prompt tuning · Camelyon17 | A-Acc: 86.31 -> 86.24 | No clear effect |
| Replace DIP fine-tuning with GAT refining (ablation) | DIP + DSP · Camelyon17 | A-Acc: 86.24 -> 90.34 | Helped |
| Replace DIP fine-tuning with GAT refining (ablation) | DIP + DSP · Camelyon17 | FTU: 77.49 -> 80.04 | Helped |
| Add style-augmented training data to GAT refining (ablation) | DIP + DSP with GAT · Camelyon17 | A-Acc: 90.34 -> 91.61 | Helped |
| Add style-augmented training data to GAT refining (ablation) | DIP + DSP with GAT · Camelyon17 | FTU: 80.04 -> 82.17 | Helped |
| Set prompt length to 1 instead of 5 (ablation) | Ours · Camelyon17 | A-Acc: 91.61 -> 90.24 | Hurt |
| Set prompt length to 3 instead of 5 (ablation) | Ours · Camelyon17 | A-Acc: 91.61 -> 91.47 | Hurt |
Surprising findings
- Sequential finetuning on the Camelyon17 stream strongly forgets earlier domains (BWT -37.13), and the joint-training upper bound is not always best on unseen domains.
Limitations
- Evaluated only on two histopathology tasks with a small number of domains (4-5).
- Training takes about 0.6h longer than other prompt-based methods on average.
Experimental setup
Datasets: Camelyon17 VGH NKI IHC NCH
Benchmarks: Camelyon17 epithelium-stroma tissue classification (VGH/NKI/IHC/NCH)
Metrics: Acc IL BWT FTU A-Acc AAMS MS
Baselines: Sequential Finetune LwF EWC SI DGR Orc-MML S-Prompt DualPrompt
Builds on: DualPrompt S-Prompt Vision Transformer graph attention network
Closest papers by meaning
- Sequential Attention-based Sampling for Histopathological Analysis (NeurIPS 2025 · similarity 0.88)
- MAPLE: Multi-scale Attribute-enhanced Prompt Learning for Few-shot Whole Slide Image Classification (NeurIPS 2025 · similarity 0.87)
- Prototype-Based Image Prompting for Weakly Supervised Histopathological Image Segmentation (CVPR 2025 · similarity 0.86)
- Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology (NeurIPS 2025 · similarity 0.85)
- xMIL: Insightful Explanations for Multiple Instance Learning in Histopathology (NeurIPS 2024 · similarity 0.85)
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.