Tissue Concepts v2 (TCv2)

• Tissue Concepts v2 (TCv2) is a supervised whole-slide image foundation model that uses weakly supervised multi-instance learning to derive slide-level representations from clinical endpoints.
• It integrates a tiny Swin Transformer v2 encoder with an 8-head attention pooling module to produce explainable, task-agnostic features for applications like cancer subtyping, survival, and mutation prediction.
• TCv2 demonstrates superior diagnostic performance with high AUC and balanced accuracy on multiple benchmarks while significantly reducing training compute compared to SSL-based methods.

Tissue Concepts v2 (TCv2) is a supervised whole-slide-image (WSI) foundation model for computational pathology that is trained end-to-end on slide-level labels using multitask learning. It extends the earlier Tissue Concepts patch-level foundation model to operate directly at the slide level, learning a shared WSI representation that supports diverse tasks such as cancer subtyping, survival prediction, and mutation status prediction. Its defining methodological move is to replace the resource-intensive self-supervised pretrain–then-aggregate pipeline with a single-stage, weakly supervised multiple-instance learning (MIL) model trained exclusively on openly available data, while retaining an attention-based pooling mechanism that provides a shared layer of explainability across tasks (Nicke et al., 8 Jul 2025).

1. Lineage and conceptual positioning

TCv2 is the slide-level extension of the original Tissue Concepts model. The earlier Tissue Concepts encoder was a supervised, multi-task, histopathology foundation model trained by combining 16 different classification, segmentation, and detection tasks on a total of 912,000 patches, with the aim of learning a reusable histopathology encoder at substantially lower training cost than self-supervised alternatives (Nicke et al., 2024). TCv2 preserves the central supervised multitask philosophy of that earlier system but relocates representation learning from the patch level to the whole-slide level (Nicke et al., 8 Jul 2025).

The conceptual distinction from prevailing WSI foundation-model pipelines is explicit. Self-supervised pathology systems typically pretrain a patch encoder with self-supervision and then build slide-level embeddings through a second aggregation stage; the patch encoder is usually not updated during that slide-level stage. TCv2 instead trains a single stack—patch encoder, attention pooler, and task heads—end-to-end on slide-level labels in a multitask supervised fashion. This means that the learned slide representation is directly shaped by clinical endpoints including subtype, mutation status, and survival, rather than by pretext tasks such as masking, contrastive views, or image–text alignment (Nicke et al., 8 Jul 2025).

A related point concerns data provenance and reproducibility. TCv2 is trained primarily on TCGA and PLCO, with TCGA and CPTAC data obtained through fully specified NCI Imaging Data Commons queries, whereas many SSL models rely on large, closed internal datasets. The paper presents this openness not as a secondary implementation detail but as part of the model’s positioning within computational pathology: efficient training of a WSI foundation model without massive private datasets, and with transparent cohort construction (Nicke et al., 8 Jul 2025).

2. Architecture and representation learning

TCv2 follows a standard MIL, or “bags-of-patches,” formulation, but it modifies that framework for end-to-end supervised multitask learning. Each WSI is treated as a bag of patches. During training, a bag consists of 64–128 randomly sampled patches from a single slide, and each patch is passed through a shared encoder based on a tiny Swin Transformer v2 initialized from the original Tissue Concepts patch model. For a slide $s$ with patch instances $\{x_j\}_{j=1}^N$, the encoder produces embeddings

$h_j = f_\theta(x_j)\in \mathbb{R}^d.$

(Nicke et al., 8 Jul 2025)

The bag of patch embeddings is reduced to a single slide-level representation by a multi-head attention pooling module inspired by attention-based MIL. TCv2 uses 8 attention heads. In ABMIL notation, one commonly writes

$$a_j = \frac{\exp\left(w^\top \tanh(V h_j^\top)\right)}{\sum_{l} \exp\left(w^\top \tanh(V h_l^\top)\right)}, \quad z_s = \sum_j a_j h_j.$$

TCv2 generalizes this with multi-head attention: each head has its own parameters and produces its own weighting, but all heads operate on the same patch embeddings from the shared Swin v2 encoder. The paper states that it does not introduce a novel attention formula; it uses an ABMIL-style multi-head attention module with 8 heads (Nicke et al., 8 Jul 2025).

A central design decision is that the attention pooling module is shared across all tasks. The same attention mechanism produces a single generic WSI embedding, and multiple task-specific heads operate on that embedding. On top of the shared $z_s$, TCv2 attaches lightweight linear heads, one per task, each consisting of a 10% dropout layer followed by a single linear layer. For task $t$, prediction is written as

$\hat{y}_s^{(t)} = g_{\phi_t}(z_s).$

This structure forces the pooling mechanism to learn a task-agnostic but broadly informative slide representation (Nicke et al., 8 Jul 2025).

The training objective is multitask and weakly supervised. If $\mathcal{T}$ denotes the task set and $\mathcal{L}_t$ the task-specific loss, the global objective can be viewed as

$$\mathcal{L} = \sum_{t \in \mathcal{T}} \lambda_t \mathcal{L}_t,$$

with implicit or equal weighting. In implementation, tasks are sampled sequentially and gradients are accumulated before optimizer steps. Shared encoder and attention parameters are updated by gradients from all tasks, whereas each task head is updated only by its own loss. There is no separate frozen-encoder stage during slide-level pretraining: gradients from slide-level labels are backpropagated through head, attention pooling, and tile encoder (Nicke et al., 8 Jul 2025).

3. Data sources, supervision, and optimization regime

TCv2 is trained on WSIs from TCGA, CPTAC, and PLCO. TCGA and CPTAC are accessed through the NCI Imaging Data Commons as DICOM whole-slide images, with CPTAC reserved for evaluation rather than pretraining. PLCO contributes 881 radical prostatectomy cases and 2999 H&E WSIs; after cleaning and exclusion of unusable endpoints, 875 cases and 2352 slides remain, split 80/20 for train and validation. Among these cohorts, PLCO is the only dataset described as non-trivially accessible (Nicke et al., 8 Jul 2025).

From these resources, 18 distinct slide-level tasks are constructed. They span three supervision families. First, organ subtype classification includes “Organ Subtype” with 22,225 WSIs, LGG/GBM subtype, kidney KICH/KIRP/KIRC subtyping, PRAD subtype, BRCA subtype, LUAD/LUSC subtype, and COAD/READ subtype. Second, overall survival prediction includes SKCM, PLCO prostate, PRAD, BRCA, and COAD/READ. Third, mutation prediction includes BRCA TP53, SPAT1, and CDH1; COAD/READ TP53 and KRAS; and LUAD/LUSC TP53 in LUSC. Each task provides one label per slide, and related tasks are split coherently to avoid leakage across shared slides (Nicke et al., 8 Jul 2025).

Training is conducted for 200 epochs on a single NVIDIA A100 GPU, with total training time of approximately 500 GPU hours for the end-to-end slide-level model. The earlier Tissue Concepts patch encoder used for initialization required 160 hours on an RTX A5000. During TCv2 training, each minibatch consists of one bag from a single WSI for a particular task; bags contain 64–128 tiles in training and always 128 tiles in validation. Training tiles are independently augmented by color shift, blur, grayscale, and geometric deformation, whereas validation is performed without augmentation (Nicke et al., 8 Jul 2025).

The resource-efficiency comparison with self-supervised WSI models is one of the paper’s most explicit claims. TCv2 training is estimated at approximately 200 kWh and 70–105 kg CO$\{x_j\}_{j=1}^N$0 for the A100 run, plus about 20 kg from the earlier TC training, for a total of roughly 90–125 kg CO$\{x_j\}_{j=1}^N$1. By comparison, the Prov-GigaPath slide encoder is reported at 3072 A100 hours for slide-level training alone, corresponding to roughly 500 kg CO$\{x_j\}_{j=1}^N$2 for that component, while CTransPath is reported at 12,000 hours on a V100. The paper therefore argues that a supervised, multitask, end-to-end MIL foundation model can be trained using a fraction, specifically 5–25%, of the compute required by leading SSL WSI models (Nicke et al., 8 Jul 2025).

4. Evaluation and comparative performance

The evaluation protocol compares TCv2 with the original Tissue Concepts patch encoder and with SSL-based WSI foundation models where published results exist. To isolate representation quality, all evaluation fine-tuning uses frozen encoders and adapts only the aggregation module (Nicke et al., 8 Jul 2025).

On CPTAC-NSCLC, an in-domain lung carcinoma subtype benchmark trained on TCGA-NSCLC and tested on CPTAC-NSCLC with 432 patients in a 40/10/50 split, TCv2 achieves $\{x_j\}_{j=1}^N$3 AUC and $\{x_j\}_{j=1}^N$4 balanced accuracy. The original TC achieves $\{x_j\}_{j=1}^N$5 AUC and $\{x_j\}_{j=1}^N$6 balanced accuracy; Prov-GigaPath reaches $\{x_j\}_{j=1}^N$7 AUC and $\{x_j\}_{j=1}^N$8 balanced accuracy; CHIEF reports $\{x_j\}_{j=1}^N$9 AUC. The paper presents this as evidence that TCv2 surpasses both CHIEF and Prov-GigaPath on this task despite using a tiny Swin v2 backbone and substantially less training data and compute (Nicke et al., 8 Jul 2025).

On PANDA, an in-domain prostate Gleason or ISUP grading benchmark with 10,616 WSIs from Karolinska and Radboud in a 50/10/40 split, TCv2 obtains $h_j = f_\theta(x_j)\in \mathbb{R}^d.$0 AUC and Cohen’s quadratic kappa $h_j = f_\theta(x_j)\in \mathbb{R}^d.$1. The original TC yields $h_j = f_\theta(x_j)\in \mathbb{R}^d.$2 AUC and $h_j = f_\theta(x_j)\in \mathbb{R}^d.$3 kappa. The contrast is used to underscore the importance of WSI-level end-to-end training: the patch-only encoder with a randomly initialized aggregation module essentially fails on this task (Nicke et al., 8 Jul 2025).

On BRACS, an out-of-domain breast cancer subtyping dataset of 547 WSIs from 189 patients with train/validation/test split 395/87/65, TCv2 reports $h_j = f_\theta(x_j)\in \mathbb{R}^d.$4 AUC and $h_j = f_\theta(x_j)\in \mathbb{R}^d.$5 balanced accuracy. The original TC reaches $h_j = f_\theta(x_j)\in \mathbb{R}^d.$6 AUC and $h_j = f_\theta(x_j)\in \mathbb{R}^d.$7 balanced accuracy; Prov-GigaPath reports $h_j = f_\theta(x_j)\in \mathbb{R}^d.$8 balanced accuracy; CHIEF reports $h_j = f_\theta(x_j)\in \mathbb{R}^d.$9 balanced accuracy. The paper characterizes this as a substantial improvement over TC and Prov-GigaPath, with balanced accuracy comparable to CHIEF (Nicke et al., 8 Jul 2025).

On the fibrosis benchmark, an out-of-domain internal dataset of 94 H&E-stained precision-cut lung slices with slide-level high-versus-low collagen area labels thresholded at the median collagen fraction, TCv2 achieves $$a_j = \frac{\exp\left(w^\top \tanh(V h_j^\top)\right)}{\sum_{l} \exp\left(w^\top \tanh(V h_l^\top)\right)}, \quad z_s = \sum_j a_j h_j.$$0 AUC and $$a_j = \frac{\exp\left(w^\top \tanh(V h_j^\top)\right)}{\sum_{l} \exp\left(w^\top \tanh(V h_l^\top)\right)}, \quad z_s = \sum_j a_j h_j.$$1 balanced accuracy. The original TC reports $$a_j = \frac{\exp\left(w^\top \tanh(V h_j^\top)\right)}{\sum_{l} \exp\left(w^\top \tanh(V h_l^\top)\right)}, \quad z_s = \sum_j a_j h_j.$$2 AUC and $$a_j = \frac{\exp\left(w^\top \tanh(V h_j^\top)\right)}{\sum_{l} \exp\left(w^\top \tanh(V h_l^\top)\right)}, \quad z_s = \sum_j a_j h_j.$$3 balanced accuracy. The paper treats this as evidence of some degree of cross-domain generalization to a target unrelated to the pretraining labels (Nicke et al., 8 Jul 2025).

Across these tasks, the main empirical pattern is consistent: TCv2 markedly improves over the original patch-only Tissue Concepts representation, especially when slide-level structure is indispensable. This supports the paper’s broader claim that supervised, multi-task, end-to-end MIL with slide-level labels yields more powerful WSI representations than patch-only supervised pretraining or standard SSL pipelines (Nicke et al., 8 Jul 2025).

5. Explainability, downstream adaptation, and generalization

The shared multi-head attention pooling module serves both representational and interpretive functions. Because each patch in a bag receives an attention weight, the model supplies a direct measure of instance-level importance for slide-level prediction. For a trained downstream task, one may sample patches, run them through the encoder and attention module, extract the attention weights $$a_j = \frac{\exp\left(w^\top \tanh(V h_j^\top)\right)}{\sum_{l} \exp\left(w^\top \tanh(V h_l^\top)\right)}, \quad z_s = \sum_j a_j h_j.$$4, and map them back to their spatial positions to generate an attention heatmap. The paper illustrates this procedure on the fibrosis task, where attention overlays on H&E precision-cut lung slices focus on collagen-rich regions despite the absence of pixel-level segmentation labels (Nicke et al., 8 Jul 2025).

This interpretability mechanism is described as shared across tasks because the same attention module is used throughout pretraining. A plausible implication is that TCv2 attempts to learn a generic notion of informative WSI regions rather than separate attention systems for each endpoint. The paper, however, is explicit that it does not provide a quantitative explainability metric such as correlation with ground-truth segmentations; the explainability evidence is qualitative (Nicke et al., 8 Jul 2025).

For downstream adaptation, the evaluation protocol freezes the Swin v2 backbone and fine-tunes only the attention-based aggregation module for 15 epochs with AdamW at learning rate $$a_j = \frac{\exp\left(w^\top \tanh(V h_j^\top)\right)}{\sum_{l} \exp\left(w^\top \tanh(V h_l^\top)\right)}, \quad z_s = \sum_j a_j h_j.$$5, selecting the best validation checkpoint. Fine-tuning is repeated four times per task and average test metrics are reported. This setup is used to demonstrate that TCv2 functions as a reusable feature extractor: competitive performance is obtained even when only the aggregator is adapted and the encoder remains fixed (Nicke et al., 8 Jul 2025).

The generalization analysis is organized around in-domain and out-of-domain transfer. In-domain tasks, such as CPTAC-NSCLC and PANDA, are closely aligned with the pretraining objectives and show strong performance. Out-of-domain tasks, such as BRACS and fibrosis, involve different sources or label types not explicitly present in pretraining. TCv2 maintains AUCs in the range $$a_j = \frac{\exp\left(w^\top \tanh(V h_j^\top)\right)}{\sum_{l} \exp\left(w^\top \tanh(V h_l^\top)\right)}, \quad z_s = \sum_j a_j h_j.$$6–$$a_j = \frac{\exp\left(w^\top \tanh(V h_j^\top)\right)}{\sum_{l} \exp\left(w^\top \tanh(V h_l^\top)\right)}, \quad z_s = \sum_j a_j h_j.$$7 on these tasks and improves balanced accuracy over the patch-only baseline. The paper does not explicitly analyze robustness to staining variation, scanner differences, or severe domain shifts beyond these cross-dataset evaluations, but it argues that the multi-organ, multi-cohort training mixture and the external-task performance suggest reasonable robustness (Nicke et al., 8 Jul 2025).

6. Practical implications, limitations, and future directions

From a practical standpoint, TCv2 is presented as a comparatively accessible WSI foundation model. It can be trained with a single A100 in approximately 500 hours, uses mostly public data, and supports straightforward downstream reuse through a frozen-encoder workflow. Because only the attention pooling module and task head need to be adapted in downstream tasks, the model lowers the barrier for institutions that lack the resources for large-scale self-supervised pretraining or access to massive private slide repositories (Nicke et al., 8 Jul 2025).

The paper also advances a methodological claim about supervision. Rather than treating slide-level labels as coarse substitutes for patch annotations, TCv2 uses subtype, mutation status, and survival labels as direct supervision for representation learning. This suggests that weak slide-level labels, if sufficiently diverse across organs and endpoints, can be used to induce a clinically oriented shared WSI space. The model’s success on multiple downstream benchmarks is presented as evidence that such supervision can act as a viable and efficient alternative to self-supervised pretraining in computational pathology (Nicke et al., 8 Jul 2025).

Several limitations are stated directly. Evaluation breadth remains limited to four external tasks. The backbone is a tiny Swin Transformer v2, and scaling experiments with larger backbones are not presented. The supervision remains weak and label-dependent; slide-level labels are easier to obtain than dense annotations but may be noisy or coarse, especially for overall survival and mutation status. The paper also does not analyze demographic or institutional biases, nor does it systematically study robustness to extreme domain shifts such as new scanners or rare stains (Nicke et al., 8 Jul 2025).

The future directions identified are correspondingly specific. They include scaling TCv2 to larger backbones, combining weakly labeled WSIs with patch-based tasks in a single multitask training regime, and expanding the evaluation benchmark suite to better quantify generalization across organs, institutions, and tasks. Within the broader foundation-model landscape of computational pathology, TCv2 is positioned as an example of how supervised multitask learning on weak slide-level labels can produce a strong, explainable, and compute-efficient WSI foundation model trained on openly available data (Nicke et al., 8 Jul 2025).