Skip to content

Temporal Inversion for Learning Interval Change in Chest X-Rays

🕒 Published (v1): 2026-04-06 09:52 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

TILA (Temporal Inversion-aware Learning and Alignment) is a training framework that uses reversed image-pair ordering as a supervisory signal to force temporal VLP models for chest X-rays to learn directional interval change, not just appearance. It introduces inversion-aware losses at pretraining, fine-tuning, and inference stages and improves progression classification and temporal embedding alignment across multiple architectures without modifying them.

Problem

Existing medical vision-language models analyze CXRs in isolation or, when paired, fail to explicitly encode the direction of temporal change. Standard progression labels (improved/stable/worsened) are noisy and do not probe whether models truly distinguish directional trajectories versus exploiting entity presence, leaving genuine temporal reasoning unverified.

Method

TILA wraps any paired-image VLP backbone with three inversion-aware components:

  1. Pretraining — Change-aware Sigmoid Loss: Extends SigLIP contrastive loss over both original and order-swapped image pairs. Unchanged pairs (ci=0) are treated as positives in both orders; changed pairs (ci=1) are treated as negatives under reversal, preventing alignment when temporal direction is reversed. Binary change labels are auto-assigned by Gemini 2.0 Flash from radiology reports.

  2. Fine-tuning — BiCE + TCL: Bidirectional Cross-Entropy (BiCE) averages forward CE loss and reversed CE loss with label-swapped targets (improved↔worsened, stable→stable). Temporal Consistency Loss (TCL) minimizes the L2 distance between forward probability vectors and the improved/worsened-swapped reversed probability vectors, enforcing probability-level symmetry.

  3. Inference — Inversion-aware scoring: Final prediction averages the forward softmax and the S-transformed (improved↔worsened swap) reversed softmax, reducing order bias without re-training.

TILA is validated on BioViL-T and ALTA backbones. Binary change labels for pretraining are generated from MIMIC-CXR using an LLM; supervised classification uses Chest ImaGenome 3-class labels on MS-CXR-T.

Key Contributions

  • Architecture-agnostic TILA framework integrating temporal inversion supervision across all training and inference stages
  • Change-aware Sigmoid Loss that differentially reweights contrastive pairs based on temporal change status
  • BiCE + TCL fine-tuning objectives enforcing label-level and probability-level inversion symmetry
  • MS-CXR-T_retrieval: a progression-aware retrieval benchmark constructed via a reproducible finding-level directional variant pipeline applicable to any annotated temporal CXR dataset
  • Unified evaluation protocol with Standard/Reversed/Combined/Consistency splits to quantify order sensitivity beyond single-metric accuracy

Results

  • Supervised (MS-CXR-T) — BioViL-T: TILA raises Standard Avg 61.1→64.1%, Consistency Avg 39.5→57.4% over SigLIP baseline; ALTA: Standard 61.7→63.6%, Consistency 42.9→54.6%
  • Zero-shot (MS-CXR-T) — BioViL-T: Standard Avg 44.0→47.2%, Consistency 24.0→30.5%; ALTA: Standard 44.0→46.4%, Consistency 18.3→28.4%
  • Private hospital cohort (supervised) — BioViL-T: Standard Avg 55.3→66.1%, Consistency 31.1→54.6%; ALTA: Standard 58.9→64.6%, Consistency 42.6→53.0%
  • TEM (temporal embedding matching) — CheXpert: BioViL-T_TILA 20.8 vs SigLIP baseline 18.5; ALTA_TILA 19.2 vs 17.8
  • Binary interval-change screening (AUC) — BioViL-T_TILA: Private 0.765 vs baseline 0.734; RexGradient 0.702 vs 0.565 (fine-tune mode)
  • Ablation on BioViL-T: Change-aware SigLIP pretraining alone raises Consistency from 39.5→45.4%; adding BiCE→53.1%; adding TCL→57.4%

Limitations

  • Inter-reader variability in interval-change labels (particularly stable vs. borderline cases) limits ground-truth reliability; TILA does not resolve annotation noise
  • Label inversion assumes improved↔worsened symmetry, which is clinically invalid for findings with asymmetric recovery trajectories (e.g., post-surgical, fibrotic changes); inference-stage inversion requires cautious interpretation for such cases
  • No unified consensus criteria exist for defining progression across different findings, complicating standardized evaluation
  • Binary change labels for pretraining are LLM-generated and not radiologist-verified, introducing potential label noise
  • Evaluated only on chest radiographs; generalizability to other imaging modalities is undemonstrated

Relevance to Foundation Models in Medicine

TILA directly addresses a core gap in medical VLPs: most foundation models encode single images and cannot model the temporal evolution central to radiological workflow. By treating temporal inversion as a self-supervisory signal, TILA is a lightweight, architecture-agnostic plug-in that can retrofit existing medical VLPs (demonstrated on BioViL-T and ALTA) with directional temporal sensitivity, without requiring architectural changes. The framework's modular design and generalization to external datasets and binary triage tasks make it practically relevant for longitudinal CXR analysis at scale. For researchers tracking medical foundation models, this work exemplifies the emerging paradigm of adding task-specific inductive biases (here, temporal directionality) to general VLPs through auxiliary losses rather than architectural redesign.