PaperlayerView exact revision on arXiv

Source-linked AI summary

Observation-Anchored Selective Assimilation for Longitudinal Tumor-State Proxy Forecasting in Post-Treatment Glioma

Yeonjae Jung, Minwoo Shin

arXiv:2609.12435v1cs.LGcs.AI

TL;DR

Post-treatment glioma MRI supplies repeated patient-specific observations, but proxy trajectories vary and direct prediction, rollout, or fixed-trend priors do not explicitly assimilate the intermediate observation. The paper uses OASA to anchor subsequent forecasting in that observed state; it matched persistence at Dice τ = 0.2 and was numerically higher at τ = 0.5.

  • Problem

    Longitudinal tumor-state proxy forecasting must address variable post-treatment appearances and trajectories, while prior direct prediction, rollout, or fixed-trend approaches do not explicitly use the observed intermediate proxy as an updated state.

    Source: §1 · pp. 1–2

  • Method

    OASA retains the observed intermediate proxy as the state anchor and selectively applies model-predicted updates through validation-selected case-level tiers and a voxel-wise soft gate.

    Source: §1 · p. 2, §2.3 · p. 4

  • Results

    OASA yielded Dice at τ = 0.2 of 0.6071 ± 0.0025 vs. 0.6070 for persistence and Dice at τ = 0.5 of 0.4269 ± 0.0079 vs. 0.3981.

    Source: §4.1 · p. 6, §5 · p. 8

  • Takeaways & Limitations

    OASA is a conservative alternative to unconditional direct prediction, comparable to persistence at Dice τ = 0.2 and numerically higher at τ = 0.5, with a small RMSE increase.

    Source: §5 · p. 8

    The study forecasts a label-derived proxy rather than clinical progression, recurrence, treatment response, or biological tumor burden, and its one-institution results from 15 held-out test triplets are descriptive.

    Source: §5 · p. 9

Abstract

from arXiv · show

Post-treatment MRI in patients with glioma provides serial observations for updating patient-specific tumor-state proxy estimates, but variable appearances and trajectories complicate forecasting. We formulate forecasting as an observation-aware digital-twin update in which an intermediate observation anchors the patient-specific state. Among 203 patients and 594 follow-up time points, a predefined no-new-treatment criterion retained 120 of 236 candidate triplets, split into 81/24/15 training/validation/test triplets at the patient level. Each time point was represented by a continuous voxel-wise tumor-state proxy map in [0,1] derived from MRI lesion labels. A SegMamba-based single-step forecaster predicted update proposals from multimodal source-state tensors. Observation-Anchored Selective Assimilation (OASA) retained the observed intermediate proxy as the state anchor and selectively applied updates through a validation-selected tiered case-level rule and voxel-wise soft gate. We compared initial-scan forecasting, rollout without assimilation, latest-observation persistence, direct prediction, OASA, OASA + calibration, and morphological dilation. Checkpoints, OASA rules, and calibration thresholds were selected using validation data only. Across three seeds on 15 held-out test triplets, OASA maintained Dice at $\tau$ = 0.2 comparable to persistence (0.6071 $\pm$ 0.0025 vs. 0.6070) while yielding numerically higher Dice at $\tau$ = 0.5 (0.4269 $\pm$ 0.0079 vs. 0.3981), with a small RMSE increase. Calibration increased Dice at $\tau$ = 0.2 to 0.6178 $\pm$ 0.0025, increased false-positive (FP) support (11,836$\rightarrow$18,663), and reduced false-negative (FN) support (22,107$\rightarrow$17,536). This reflects near-threshold support calibration rather than improved biological predictive capability. Code is publicly available at https://github.com/jsudg436/longitudinal-proxy-forecasting.

1 Introduction

Post-treatment tumor-state proxy trajectories can be stable, increasing, decreasing, or non-monotonic, while direct prediction, rollout, and fixed-trend priors do not explicitly use the observed intermediate proxy as the updated patient-specific state [2] [21]. The paper frames forecasting as an observation-aware digital-twin update and introduces OASA to anchor and selectively update that state.

1 Introduction

  • Proxy trajectories may be stable, increasing, decreasing, or non-monotonic, limiting a single fixed extrapolation rule.The forecast target is a continuous voxel-wise proxy map derived from MRI lesion labels.Source: §1 · pp. 1–2
  • Direct t0-to-t2 prediction, rollout, and fixed growth or trend priors do not explicitly use the observed t1 proxy as the updated patient-specific state.Prior work on longitudinal prediction, biophysical growth, reaction–diffusion modeling, and MRI synthesis motivates patient-specific forecasting [2] [21].Source: §1 · p. 2
  • OASA retains the observed intermediate proxy as the state anchor and selectively applies model-predicted updates through tiered case-level and voxel-wise soft gates.The single-step forecaster uses a SegMamba-based 3D encoder–decoder; the contribution is the assimilation formulation rather than a claim of backbone superiority.Source: §1 · p. 2
  • The contributions include patient-level evaluation using continuous proxy error and proxy-support overlap at τ = 0.2 and τ = 0.5, plus descriptive characterization of heterogeneous trajectories.No checkpoint, OASA rule, calibration threshold, or final seed was selected based on held-out test performance.Source: §1 · p. 2

2 Method

The method forecasts continuous tumor-state proxy maps from multimodal source states, then uses the observed intermediate proxy to anchor a selectively gated update. A validation-selected calibration can adjust near-threshold support.

2.1 Tumor-State Proxy Forecasting

  • Each time point is represented by a [0,1] proxy map derived from glioma labels, and each forecast uses a seven-channel source-state tensor.The tensor combines four MRI channels, the source proxy, a cavity map, and a normalized time-interval channel.Source: §2.1 · pp. 2–3

2.2 Single-Step Forecaster

  • A SegMamba-based forecaster predicts normalized velocity and voxel-wise change probability, which combine with the source proxy and interval to form a clipped destination forecast.The effective velocity is probability-weighted, and the training objective combines change-weighted L1, Dice, velocity, change-logit, and stable-delta losses.Source: §2.2 · pp. 3–4

2.3 Observation-Anchored Selective Assimilation

  • OASA anchors the future proxy at observed t1, using innovation MAE and predicted-change burden to select an update tier before voxel-wise gating.The soft gate weights the already probability-weighted proposal; persistence applies no update.Source: §2.3 · p. 4
  • The tier rule uses predicted-change fraction B and innovation MAE I, with validation-selected thresholds and tier-specific update factors; persistence applies no update.The tier labels denote rule branches, not calibrated confidence levels.Source: §2.3 · p. 4
  • The final OASA update is the selected tier factor times a soft voxel gate applied to the proposed update, added to observed t1 and clipped to [0,1].The gate applies additional change-score weighting to an already probability-weighted proposal.Source: §2.3 · p. 4

2.4 Near-Threshold Calibration

  • Calibration raises eligible voxels from a validation-selected lower threshold of 0.075 to 0.201, just above the evaluation threshold of 0.2.The paper characterizes this as near-threshold support calibration, not improved biological predictive capability.Source: §2.4 · p. 5

3 Experiments

The experiments use a treatment-filtered cohort with patient-level splits, evaluate continuous proxy error and support overlap, and compare forecasting methods under validation-only selection.

3.1 Data, Splits, and Metrics

  • From 203 patients, the predefined no-new-treatment filter retained 120 of 236 candidate triplets from 72 patients, split 81/24/15 by patient.The retained set was used for modeling and trajectory characterization.Source: §3.1 · p. 5
  • RMSE measured continuous proxy error, while Dice at τ = 0.2 and τ = 0.5 measured broader and more restrictive proxy-support overlap.Metrics covered all voxels in the centered 128^3 evaluation patch without an additional brain mask; paired uncertainty used 10,000 triplet-level bootstrap resamples after averaging across seeds.Source: §3.1 · p. 5

3.2 Baselines and Validation Protocol

  • Comparisons included initial-scan forecasting, rollout, persistence, direct prediction, OASA, OASA + calibration, and validation-selected radius-1 morphological dilation.Initial forecasting predicted t2 from t0; rollout propagated a t0-to-t1 prediction, persistence copied observed t1, and direct prediction forecast t2 from observed t1 without OASA gating.Source: §3.2 · p. 6
  • Checkpoints, OASA rules, calibration thresholds, and baseline parameters were selected on validation data, then applied once to 15 held-out test triplets without test-based method or seed selection.The fixed patient-level split was shared across three seeds.Source: §3.2 · p. 6

4 Results

OASA retained the observed intermediate proxy as its anchor and selectively applied forecast updates. It matched persistence at the lower support threshold and had numerically higher overlap at the higher threshold, while trajectories and support patterns varied.

4.1 Main Comparison

  • 0.4269 ± 0.0079 vs. 0.3981 for OASA vs. persistence at Dice τ = 0.5; at τ = 0.2, scores were 0.6071 ± 0.0025 vs. 0.6070.OASA had a small RMSE increase; Fig. 2 presents the performance trade-off and descriptive support changes.Source: §4.1 · p. 6, Figure 2 · p. 7
  • 0.6178 ± 0.0025 was OASA + calibration’s Dice at τ = 0.2; calibration increased FP@0.2 from 11,836 to 18,663 and reduced FN@0.2 from 22,107 to 17,536.Calibration did not change aggregate Dice at τ = 0.5; morphological dilation had lower Dice at τ = 0.5 and a higher FP voxel count at τ = 0.2 than OASA.Source: §4.1 · p. 7
  • 0.1493 (95% CI 0.0291–0.3055) was OASA’s Dice τ = 0.5 increase over direct prediction, alongside a 0.0202 (95% CI 0.0113–0.0294) RMSE reduction.CIs for OASA versus persistence and morphological dilation included zero for Dice metrics.Source: §4.1 · p. 7

4.2 Trade-off, Trajectories, and Qualitative Update

  • Non-monotonic trajectories were the largest category (54/120, 45.0%), followed by mixed/mild (26/120, 21.7%), increase-like (24/120, 20.0%), stable-like (10/120, 8.3%), and decrease-like (6/120, 5.0%).The descriptive analysis motivates observation-aware updating rather than a single monotonic extrapolation rule.Source: §4.2 · p. 7, Figure 2 · p. 7
  • Fig. 3 shows different support patterns for direct prediction, OASA, and OASA + calibration against the same future target proxy; OASA uses a conservative gated update from the observed intermediate anchor.The displayed case was selected for visualization after model, checkpoint, rule, threshold, and seed choices were fixed.Source: §4.2 · pp. 7–8, Figure 3 · p. 8

5 Discussion and Conclusion

OASA was comparable to persistence at Dice τ = 0.2 and numerically higher at τ = 0.5, while calibration shifted near-threshold support rather than improving biological prediction. The findings concern label-derived proxy forecasting, not clinical outcomes or untreated biological trajectories.

5 Discussion and Conclusion

  • OASA matched persistence at Dice τ = 0.2 and was numerically higher at τ = 0.5, while direct prediction showed greater across-seed variability.Morphological dilation had higher Dice and FP voxel count at τ = 0.2, but lower Dice at τ = 0.5 than OASA.Source: §5 · p. 8
  • OASA + calibration reduced FN and increased Dice at τ = 0.2, but increased FP; it is a near-threshold support adjustment, not improved biological prediction.By construction, calibration left Dice at τ = 0.5 unchanged.Source: §5 · pp. 8–9
  • No-new-treatment proxy trajectories were frequently non-monotonic, but should not be considered untreated biological trajectories; SegMamba is the backbone, not the contribution.The study motivates growth and reaction–diffusion extrapolation but neither reproduces those frameworks nor claims superiority.Source: §5 · p. 9
  • The study forecasts a label-derived proxy rather than clinical progression, recurrence, treatment response, or biological tumor burden; results from one institution and 15 test triplets are descriptive.The authors call for larger multi-institution cohorts and clinical endpoints.Source: §5 · p. 9
Loading 2609.12435v1…