Source-linked AI summary
An integrated diffusion-weighted imaging processing and interpretation platform for MR-guided radiotherapy
Yunxiang Li, Yan Dai, Yen-Peng Liao, Jie Deng, Jill B De Vis, You Zhang
TL;DR
MR-Linac diffusion-weighted imaging needs reliable quantitative processing and clinically interpretable synthesis of scattered, sometimes contradictory evidence. This paper presents an integrated web-based platform linking DWI processing with traceable RAG interpretation, whose reports achieved a pooled score of 4.65 ± 0.80 and 93% of ratings at good or above.
Problem
Reliable diffusion parameters still require interpretation across literature that is scattered and often contradictory.
Method
The study develops and evaluates an integrated web-based platform combining MR-Linac DWI processing, longitudinal analysis, and traceable literature-grounded RAG interpretation.
Results
4.65 ± 0.80 was the overall report score on a 1–5 scale, with 93% of ratings at good or above.
Takeaways & Limitations
The platform provides a usable integration of MR-Linac DWI processing with clinically evaluated, traceable interpretation.
Takeaways & Limitations
Residual risk remains because the LLM-based interpreter can exhibit behaviors requiring verification against reference sources.
Abstract
from arXiv · showhide
Background: Magnetic resonance imaging-guided linear accelerators (MR-Linacs) allow diffusion-weighted imaging (DWI) to be acquired at every treatment fraction, but converting these low-signal-to-noise-ratio acquisitions into clinical decisions requires both reliable quantitative processing and an interpretation that reconciles a scattered and often contradictory literature. Purpose: To describe and evaluate an integrated, web-based platform that carries raw MR-Linac DWI to a structured, literature-grounded clinical interpretation, and to assess its retrieval-augmented generation (RAG) interpretation module by independent expert rating. Methods: The platform couples a deep-learning processing pipeline, comprising distortion correction, denoising, and intravoxel incoherent motion (IVIM)/apparent diffusion coefficient (ADC) fitting, with longitudinal region-of-interest analysis and a RAG interpretation agent. The agent reasons over a two-layer knowledge base of curated publications (a structured catalog index plus line-indexed full text), delegates arithmetic to deterministic tools, and is designed to trace each statement to a source document, section, and line range. One medical physicist and one physician independently rated the agent's reports for nine longitudinal glioblastoma cases on a 1-5 scale across three metrics: clinical-reasoning soundness, literature-citation quality, and overall clinical utility. Results: Across 54 ratings, the pooled mean was 4.65 +/- 0.80, with 93% of ratings >= 4; metric means were 4.6 (reasoning), 4.5 (citation), and 4.8 (utility), and raters agreed within one point on 85% of paired ratings. Conclusions: A single platform can integrate MR-Linac DWI post-processing with traceable, expert-evaluated clinical interpretation, while highlighting the safeguards needed to verify LLM-generated reasoning in radiation oncology.
1 Introduction
MR-Linac DWI enables dense, fraction-by-fraction tumor monitoring, but low-SNR acquisition and fragmented, contradictory evidence impede reliable clinical interpretation. This work introduces an integrated, auditable platform linking quantitative processing with literature-anchored RAG reporting and evaluates its reports in nine longitudinal GBM cases.
- Clinical opportunity: Per-fraction MR-Linac imaging enables longitudinal DWI sampling that conventional diagnostic follow-up cannot match.DWI and IVIM parameters can, in principle, provide an early, biologically specific measure of treatment response.
- Clinical opportunity: ADC reflects tissue cellularity, while IVIM separates true diffusion (Dt), pseudo-diffusion (Dp), and perfusion fraction (fp).This decomposition is intended to disentangle cellular and microvascular contributions.
- Technical and interpretive barriers: Low SNR and geometric distortion in pragmatic MR-Linac DWI destabilize quantitative fitting, particularly for perfusion-sensitive Dp and fp.The study addresses this technical problem with distortion correction, denoising, and IVIM parameter estimation components.
- Technical and interpretive barriers: Clinical interpretation remains difficult because literature thresholds and parameter-outcome relationships vary across acquisition, hardware, tumor, and treatment conditions.Dt, Dp, and fp may also move in opposite directions, with no validated framework for resolving such discordance.
- Study contribution: The platform carries raw MR-Linac DWI through processing, registration, and longitudinal ROI trajectories to a RAG agent whose report statements are anchored to source-literature passages.Reports were evaluated on nine longitudinal GBM cases by a medical physicist and a physician across reasoning, citation, and utility metrics.
2 Methods
The platform integrates DICOM handling, deep-learning DWI processing, longitudinal ROI analysis, and a traceable interpretation agent in one browser-based workflow. Its safeguards include local image processing, auditable numerical inputs, passage-level citations, and expert-verifiable reasoning.
- Data governance: Image processing runs locally on GPU-accelerated institutional hardware, while only derived numerical summaries are sent to the HIPAA-compliant language-model deployment.No image data leaves the local environment.
- Integrated platform: The single application carries studies from DICOM import through quantitative maps to clinical reports without leaving the browser or writing code.Its user-facing capabilities include ADC/IVIM processing, longitudinal ROI analysis, clinical interpretation, and DICOM export.
- DWI processing: 0.919 mean Dice coefficient was achieved for brain glioblastoma distortion correction, exceeding established diffeomorphic and learning-based baselines.The method uses cross-modality landmark matching with a B-spline implicit neural representation to produce smooth, anatomically plausible deformations.
- DWI processing: 40% of the original noise standard deviation was reached while preserving the signal mean, with a structural similarity index of 0.933 after denoising.The self-supervised band-limited implicit neural representation uses multi-b-value signal decay and cross-modality structural consistency as constraints.
- Interpretation agent: The interpretation agent receives structured, auditable numerical changes and combines two-stage retrieval, tool-assisted calculations, and passage-level citation.Derived statements are propagated to exact cited paper lines, allowing readers to verify each claim and making the system auditable.
3 Results
The platform processed MR-Linac DWI studies end-to-end and generated longitudinal IVIM/ADC analyses with traceable clinical interpretations. Expert ratings were high overall, although several reports showed grounding-related failure modes.
- Platform output: The platform processed MR-Linac studies from raw DICOM to registered longitudinal IVIM/ADC maps and clinical interpretation without manual scripting.Native-resolution fitting kept per-study processing tractable, while ROI-based parameter trajectories supported direct review and agent input.
- Representative case: The agent interpreted early diffusivity increases as treatment response and later Dt decline with preserved or rising fp as warranting attention for possible tumor regrowth.The report explicitly highlighted discordant decreasing diffusivity and stable or increasing perfusion, citing matched literature by section and line range.
- Expert evaluation: 4.65 ± 0.80 was the pooled mean across 54 ratings, with 93% of ratings ≥4; metric means were 4.61 ± 0.92 for clinical-reasoning soundness, 4.50 ± 0.99 for literature-citation quality, and 4.83 ± 0.38 for overall clinical utility.Reviewers agreed within one point on 85% of paired ratings, with a mean absolute difference of 0.63 points.
- Limitations: Grounding failures included unsupported relevance claims, uncited supplementary analyses, inferred acquisition timelines, and unexplained high-b-value diffusivity measures.These cases received lower reasoning or citation scores and showed that passage tracing did not always achieve its intended purpose.
4 Discussion
The platform unifies reliable MR-Linac DWI processing with literature-grounded, traceable interpretation and received favorable expert ratings. Discussion also identifies grounding failure modes, safeguards, broader scope, and limitations preventing deployment-ready generalization.
- Integrated platform: The platform combines validated DWI processing, longitudinal ROI analysis, and citation-traced RAG interpretation in one auditable workflow.It takes raw DICOM through distortion correction, denoising, and IVIM/ADC fitting before structured interpretation.
- Design rationale: Traceable citations, deterministic calculations, and tumor-aware retrieval supported verification, reproducible numerical interpretation, and clinically relevant evidence grounding.Citation quality declined when claims lacked matched evidence, showing that traceability depends on consistent grounding.
- Limitations and safeguards: The agent inferred scan timing, used mismatched GBM cohorts, and suggested unsupported analyses, exposing residual risks in LLM-based interpretation.Proposed safeguards include verified metadata, compatible cohorts, citations for analyses, constrained parameters, and reduced confidence for atypical cases.
- Scope and study limitations: The architecture extends beyond longitudinal GBM and is potentially generalizable across organs, while the evaluation remains single-institution, pilot-scale, and outcome-unvalidated.The knowledge base spans head-and-neck, prostate, and pancreatic disease, but the study used nine cases, two raters, and one LLM configuration.
5 Conclusions
The study presents a verifiable web-based platform that integrates MR-Linac DWI processing with literature-grounded, traceable clinical interpretation. Independent evaluation found the reports clinically usable while identifying correctable limitations and supporting broader multiparametric IVIM treatment-response monitoring.
- 5 Conclusions: The platform processes raw MR-Linac DICOM diffusion-weighted imaging into structured, literature-grounded clinical interpretation.It combines validated deep-learning distortion correction, denoising, and IVIM/ADC fitting with retrieval-augmented interpretation.
- 5 Conclusions: The retrieval-augmented interpretation agent links its statements to specific passages in the source literature.This traceable design supports auditable LLM-based reasoning in radiation oncology.
- 5 Conclusions: The evaluation identified primarily inferred acquisition metadata and imperfectly matched supporting evidence as remaining limitations.The traceable platform localized these limitations to specific, correctable statements.
- 5 Conclusions: Integrating quantitative processing and literature-grounded interpretation in one verifiable platform supports broader multiparametric IVIM treatment-response monitoring beyond specialized centers.The authors characterize this integration as a step toward wider use of multiparametric IVIM.
Ethics Approval
All datasets were retrospectively collected from an approved UT Southwestern Medical Center study under umbrella IRB protocol 082013-008; the work was not a clinical trial.
- All datasets were retrospectively collected from an approved study at UT Southwestern Medical Center.
- The analysis was retrospective rather than a clinical trial, and no clinical trial ID number was available.
Data Availability Statement
The study data cannot be publicly released because they contain sensitive personal information, but supporting data are available from the authors upon reasonable request.
- Study data are not publicly available because they contain sensitive personal information, but they can be obtained from the authors upon reasonable request.