CAMP: Continuous and Adaptive Learning Model in Pathology

TL;DR

CAMP introduces a continual, adaptive, text-generating framework for pathology image classification that couples a shared visual encoder and language decoder with task-specific adaptors stored in a dedicated adaptor storage. By freezing foundational components and updating only lightweight adapters via two loss terms, CAMP achieves strong accuracy across patch- and slide-level tasks while dramatically reducing computation and storage needs. The approach demonstrates robust performance gains over foundation models and fully fine-tuned baselines across 22 datasets (17 patch-level, 5 slide-level) and emphasizes the importance of pathology-specific priors and interpretability through attention heatmaps. This work suggests a scalable path toward a universal, digitized pathology workflow with efficient continual learning across diverse diagnostic tasks.

Abstract

There exist numerous diagnostic tasks in pathology. Conventional computational pathology formulates and tackles them as independent and individual image classification problems, thereby resulting in computational inefficiency and high costs. To address the challenges, we propose a generic, unified, and universal framework, called a continuous and adaptive learning model in pathology (CAMP), for pathology image classification. CAMP is a generative, efficient, and adaptive classification model that can continuously adapt to any classification task by leveraging pathology-specific prior knowledge and learning taskspecific knowledge with minimal computational cost and without forgetting the knowledge from the existing tasks. We evaluated CAMP on 22 datasets, including 1,171,526 patches and 11,811 pathology slides, across 17 classification tasks. CAMP achieves state-of-theart classification performance on a wide range of datasets and tasks at both patch- and slide-levels and reduces up to 94% of computation time and 85% of storage memory in comparison to the conventional classification models. Our results demonstrate that CAMP can offer a fundamental transformation in pathology image classification, paving the way for the fully digitized and computerized pathology practice.

Figures (9)

• Figure 1: Overview of CAMP for patch-level classification. $\bold{a}$) For each patch classification task, the image-text prompt input and text ground truth are generated. The patch query generation is generated by a pre-trained visual encoder and a pre-trained text decoder. $\bold{b}$) During training, $\mathcal{L}_{\mathcal{S}}$ is used for optimizing adaptors, whereas $\mathcal{L}_{\mathcal{K}}$ is utilized for updating a key. This process only updates the training task and preserves the knowledge of previously learned tasks. $\bold{c}$) During inference, a query is generated based on an input to retrieve the most suitable adaptors. After being integrated with the adaptors, CAMP generates a textual prediction.
• Figure 2: Overview of CAMP for slide-level classification. $\bold{a}$) For each slide classification task, the image-text prompt input and text ground truth are generated. The slide query generation is produced by a pre-trained visual encoder, a pre-trained text decoder, and a non-parametric aggregator. $\bold{b}$) Similar to patch-level, $\mathcal{L}_{\mathcal{S}}$ and $\mathcal{L}_{\mathcal{K}}$ are used for optimizing adaptors and a key during training a current task. A visual encoder is frozen in this process. $\bold{c}$) The slide-level inference is similar to patch-level, except for the adaptors. Note that the aggregator (blue) in the generative model is parametric, which is different from the non-parametric aggregator (grey) in the query generation procedure.
• Figure 3: Datasets utilized for experiments. $\bold{a}$) 1,171,526 patches and 11,811 slides from 8 organs are curated for comprehensive experiments. $\bold{b}$) Class distribution of 6 slide-level datasets from 3 organs. $\bold{c}$) Class distribution of 17 patch-level datasets from 8 organs.
• Figure 4: Performance of foundation models when integrated into CAMP. $\bold{a}$) CAMP increases the performance of CTransPath ctranspath, UNI uni, and Phikon phikon on a wide range of datasets, on both patch- and slide-level classification. $\bold{b}$-$\bold{c}$) The detailed comparison in patch and slide datasets, respectively. The percentages show the ratios of change in the F1 score.
• Figure 5: Comparison between the performance of efficiently finetuned CAMP and fully finetuned models. $\bold{a}$) The three configurations under consideration are task-specific, task-agnostic, and task-agnostic generative classification. $\bold{b}, \bold{d}$) CAMP performs better than other considered methods in 16/17 patch-level datasets. The numbers in the bar show the gap between CAMP and the second-best competitors. $\bold{c}$) Comparison between CAMP and competitors in terms of the computation time and memory consumption during training and inference.