PaperlayerView exact revision on arXiv

Source-linked AI summary

Benchmarking Data-Free Continual Model Merging for Whole-Slide Image Survival Analysis

Vu Minh Tran, Doanh C. Bui, Maï K. Nguyen, Khang Nguyen

arXiv:2610.32207v1cs.CV

TL;DR

WSI survival analysis needs to integrate knowledge from sequential cancer cohorts, but the relative behavior of data-free merging strategies in this setting is insufficiently understood. The paper benchmarks six strategies across cohorts and task orders under a standardized pipeline; results show that no method consistently dominates across evaluation criteria.

  • Problem

    The relative behavior of data-free merging strategies for WSI survival prediction remains insufficiently understood, despite the need to integrate sequential cohort knowledge while retaining prior learning.

    Source: §1 · pp. 1–2

  • Method

    The benchmark standardizes task-specific fine-tuning, sequential merging, and TASK-IL and CLASS-IL inference, reusing the same experts across task orders.

    Source: §1 · p. 2

  • Results

    No merging strategy consistently dominates across predictive performance, knowledge retention, and backward transfer.

    Source: §6 · p. 13

  • Takeaways & Limitations

    The benchmark provides a reproducible foundation for systematic evaluation and future development of continual model merging for WSI survival analysis.

    Source: §6 · p. 13

    Merging does not access training data from previously observed tasks, while task-specific survival heads are retained for inference.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

Continual survival analysis on whole-slide images (WSIs) requires models to integrate knowledge from sequentially arriving cancer cohorts while preserving previously learned information. Model merging offers a data-free alternative by combining independently fine-tuned task experts without retaining previous training data. However, the relative behavior of different merging strategies for WSI-based survival prediction remains insufficiently understood. In this work, we establish an end-to-end benchmark for data-free continual model merging, covering task-specific fine-tuning, sequential merging, and inference under both TASK-IL and CLASS-IL settings. We systematically evaluate six representative merging strategies across four TCGA cancer cohorts and three task orders using C-index, IPCW C-index, Forgetting, and Backward Transfer under a controlled protocol. Results show that no merging strategy consistently dominates across all evaluation criteria, while censoring-aware evaluation can substantially alter the relative ranking of the compared methods. This benchmark provides a unified foundation for developing and evaluating more effective continual model merging methods for WSI-based survival analysis.

1 Introduction

WSI survival models are commonly built separately for each cancer cohort, making continual learning costly, while data-free merging offers an alternative whose behavior in survival analysis remains insufficiently understood. This work benchmarks a standardized pipeline spanning task-specific fine-tuning, sequential merging, and TASK-IL and CLASS-IL inference.

1 Introduction

  • Separate cohort-specific WSI survival models make continual deployment costly, motivating sequential learning that retains prior knowledge.Existing continual approaches commonly use regularization or rehearsal, with rehearsal requiring stored samples from previous tasks.Source: §1 · p. 1
  • Different data-free merging strategies remain insufficiently compared for censored survival prediction and inference with known or unknown task identity.A common training, merging, and inference protocol is needed for fair comparison.Source: §1 · p. 2
  • The benchmark standardizes shared-pretrained task-expert fine-tuning, sequential merging, and inference under both TASK-IL and CLASS-IL.The same independently trained experts are reused across task orders to isolate the merging strategy from task-specific training.Source: §1 · p. 2
  • It evaluates six merging strategies across four TCGA cohorts and three task orders using survival and continual-learning metrics under a controlled protocol.The contributions specify conventional and censoring-aware survival metrics alongside continual-learning measures.Source: §1 · p. 2

2 Related Work

Related work traces model merging from fixed-model parameter combinations to continual methods that integrate arriving task experts. Continual merging has been studied predominantly for classification, leaving its behavior in WSI survival analysis insufficiently explored.

Pathology Foundation Models

  • Pathology foundation models, including TITAN, provide transferable slide-level representations and a strong initialization for WSI survival prediction and continual adaptation.The cited background describes a progression from patch-level representations to vision-language and slide-level modeling.Source: Pathology Foundation Models · p. 2

Model Merging

  • Model merging combines independently fine-tuned models without original training data, progressing from parameter averaging to task-vector and conflict-aware methods.The cited examples include Model Soups, Task Arithmetic, and TIES-Merging.Source: Model Merging · pp. 2–3

Continual Model Merging

  • Continual merging integrates arriving task models without revisiting prior task data, but existing evaluations have focused predominantly on classification.OPCM, DOP, and NUFILT are cited as methods for this setting; WSI survival also requires attention to censored outcomes and survival-specific evaluation.Source: Continual Model Merging · p. 3

3 Methodology

The methodology defines continual merging of WSI survival experts, then benchmarks a pipeline of task-specific fine-tuning, sequential merging, and inference under TASK-IL and CLASS-IL. It merges aggregator parameters while retaining task-specific survival heads for inference.

3.1 Problem formulation

  • The formulation treats each distinct cancer cohort as a survival task and merges aggregator parameters in the specified arrival order.Only aggregator parameters are merged; the task-specific survival heads are retained for inference.Source: §3.1 · p. 3

3.2 Overview of the Benchmark Pipeline

  • The benchmark pipeline comprises independent task-specific fine-tuning, sequential continual merging, and survival inference in both learning settings.The pipeline overview is shown in Fig. 1.Source: §3.2 · p. 3, Figure 1 · p. 4

3.3 Stage 1: Offline Task-specific Fine-tuning

  • Each survival task is independently fine-tuned from the same pretrained TITAN aggregator, producing task-specific aggregator and Cox survival-head parameters.The aggregator and head are jointly optimized using the Cox partial likelihood loss.Source: §3.3 · p. 4

3.4 Stage 2: Continual Model Merging Protocol

  • Six merging strategies progressively integrate task-specific aggregators; survival heads remain unchanged, and previously observed task data are not accessed during merging.The strategies include Weight Averaging, Task Arithmetic, TIES-Merging, DOP, NUFILT, and OPCM.Source: §3.1 · p. 3, §3.4 · p. 5, Weight Averaging (WA) · p. 5

3.5 Stage 3: Inference Protocols

  • The merged aggregator is evaluated under TASK-IL with the known task head and under CLASS-IL with task identity unavailable.CLASS-IL uses Voting-Expert Aggregation with task-level prompt embeddings.Source: CLASS-IL · p. 7, TASK-IL · p. 7

4 Evaluation Protocol

The protocol evaluates continual survival prediction across four TCGA cohorts, using shared task experts across multiple arrival orders and measuring both survival performance and continual-learning behavior.

Datasets

  • Four TCGA cohorts—BLCA, UCEC, LUAD, and BRCA—form the continual tasks, with patient-level survival time and censoring as targets.Five-fold patient-level cross-validation keeps slides from the same patient within a single fold.Source: Datasets · p. 7, Table 1 · p. 8

Continual Task Orders

  • The same independently fine-tuned task experts are reused across three arrival sequences, isolating task-order effects on merging performance.The sequences vary the ordering of BLCA, UCEC, LUAD, and BRCA.Source: Continual Task Orders · p. 8

Metrics

  • Performance and continual-learning behavior are assessed with C-index, IPCW C-index, Forgetting, and Backward Transfer.Higher IPCW C-index indicates better censoring-adjusted concordance; lower forgetting indicates better retention, while positive BWT indicates improved earlier-task performance.Source: Metrics · pp. 8–9
  • The IPCW C-index weights observed events using a censoring-survival estimate from the training fold, lower-bounded at 10^-12 for numerical stability.The truncation horizon is the maximum observed follow-up time in the training fold.Source: Metrics · p. 8

Training and Inference

  • Task-specific TITAN aggregators are fine-tuned jointly with Cox survival heads for 20 epochs using 1,000 randomly sampled patches per WSI.Features come from CONCH v1.5 at 10× magnification; experiments use an NVIDIA A100 GPU.Source: Implementation Details · p. 9

Evaluation

  • The final merging-step results average performance across three task orders, with means and standard deviations computed over five patient-level folds.The protocol highlights best and second-best results for comparison.Source: Table 2 · p. 10

5 Results

Across the evaluated settings, no merging strategy consistently leads on predictive performance, forgetting, and backward transfer. Censoring-aware scoring changes method rankings, while task-order variation is generally small within the tested sequences.

5.1 Overall Benchmark Performance

  • No method dominates across predictive performance, forgetting, and backward transfer: conventional C-index leaders differ from IPCW leaders, and FGT/BWT leaders also vary.In CLASS-IL, NUFILT leads conventional C-index and forgetting, while TIES leads IPCW C-index and forgetting; WA and OPCM lead BWT under conventional and IPCW evaluation, respectively.Source: §5.1 · p. 10, Table 3 · p. 11

5.2 Impact of Censoring-aware Evaluation

  • Under TASK-IL, C-index favors NUFILT (0.6543) but IPCW C-index favors TA (0.6147); under CLASS-IL, NUFILT leads C-index (0.6571) while TIES leads IPCW C-index (0.6152).NUFILT's conventional C-index falls to 0.6078 under TASK-IL IPCW evaluation and 0.6089 under CLASS-IL IPCW evaluation.Source: Table 3 · p. 11, §5.2 · p. 12

5.3 Robustness to Task Order

  • Across the three task sequences, C-index and IPCW C-index ranges remain below 0.01, while TIES varies by only 0.0002 and 0.0001 in TASK-IL and CLASS-IL IPCW C-index, respectively.NUFILT's strong conventional C-index does not imply equally low variation under censoring-aware evaluation.Source: Table 4 · p. 11, §5.3 · p. 12

5.4 Discussion

  • The authors suggest differing update-combination mechanisms may produce trade-offs between shared-knowledge preservation and task-specific information; the benchmark covers one TITAN backbone, fixed patch sampling, and four TCGA cohorts.Cross-center domain shifts and imbalanced task sequences remain unexplored.Source: §5.4 · p. 12

6 Conclusion

The conclusion presents a unified benchmark for data-free continual merging in WSI survival analysis and finds complementary trade-offs rather than a consistently dominant strategy.

6 Conclusion

  • No merging strategy consistently dominates across predictive performance, forgetting, backward transfer, and censoring-aware discrimination.The benchmark reveals complementary trade-offs among merging mechanisms.Source: §6 · p. 13
  • The benchmark spans independent task-specific fine-tuning, sequential merging, and survival inference under TASK-IL and CLASS-IL.It evaluates six representative strategies across multiple cancer cohorts and task orders.Source: §6 · p. 13
  • The unified protocol provides a reproducible foundation for evaluating and developing continual model merging methods in computational pathology.Source: §6 · p. 13

Related papers

Loading 2610.32207v1…