Source-linked AI summary
An Agentic AI Workflow for Liver Lesion Segmentation and Structured Radiology Reporting
Fatma Ahmed Mansour, Rahatara Ferdousi, Manar Fathy Elshenawy
TL;DR
Integrating segmentation-derived measurements into structured reports while preserving traceability remains a challenge, particularly because LLMs may produce numerical findings unsupported by image evidence. The framework combines liver and lesion segmentation, deterministic extraction of quantitative measurements, and LLM-based composition of structured reports with numerical consistency checks. The segmentation stage achieved liver Dice coefficient 0.947 and lesion Dice score 0.63; lesion sensitivity was 75.6% and PPV was 68.9%.
Problem
Integrating segmentation-derived measurements into structured reports while preserving traceability remains a challenge, particularly because LLMs may produce numerical findings unsupported by image evidence.
Source: §I · p. 1Method
The framework combines liver and lesion segmentation, deterministic extraction of quantitative measurements, and LLM-based composition of structured reports with numerical consistency checks.
Source: §I · p. 1, §III · pp. 2–3Results
The segmentation stage achieved liver Dice coefficient 0.947 and lesion Dice score 0.63; lesion sensitivity was 75.6% and PPV was 68.9%.
Source: §V · p. 3Takeaways & Limitations
The framework separates image-derived measurements from LLM report composition and provides a traceable approach to organizing quantitative imaging findings.
Source: §VI · p. 4, §VII · p. 4Evaluation is limited to 51 LiTS cases, with published results used as reference rather than direct comparison; measurement accuracy, reporting consistency, and clinical usability remain for future evaluation.
Source: §VI · p. 4
Abstract
from arXiv · showhide
Despite advances in automated liver lesion segmentation, a gap remains between image-derived predictions and clinically interpretable radiology reports. Bridging this gap requires quantitative measurements that are both reproducible and traceable. We present an evidence-grounded agentic AI framework that connects liver and lesion segmentation with deterministic quantification and structured reporting. TotalSegmentator-based nnU-Net models delineate the liver, Couinaud segments, and lesions, while computational tools extract volumes, dimensions, counts, and anatomical locations. A large language model (LLM) organizes these measurements into a predefined reporting schema, with numerical consistency checks against source measurements. On 51 LiTS cases, the segmentation stage achieved a liver Dice score of 0.947, lesion sensitivity of 75.6%, and positive predictive value of 68.9%. Exploratory screening of 25,494 MERLIN reports identified 10,208 (40.0%) with liver-related abnormalities. By connecting quantitative imaging evidence with agentic report generation, this work takes a step toward human-centered medical AI that prioritizes traceability and clinical relevance.
I. INTRODUCTION
Quantitative liver-lesion assessment matters for treatment planning and monitoring, but manual measurement is time-consuming and observer-variable, while traceability from image predictions to reports remains a challenge. The proposed workflow connects segmentation-derived measurements to structured reporting, with deterministic computation and checks intended to keep reported values traceable to their sources.
I. INTRODUCTION
- The framework integrates liver and lesion segmentation, deterministic quantitative analysis, and structured radiology reporting in an evidence-grounded workflow.Its contribution is the integration and traceability, rather than a new segmentation model or reporting algorithm.Source: §I · p. 1, §II · p. 2
- Pretrained TotalSegmentator nnU-Net models generate masks for liver parenchyma, Couinaud segments, and focal hepatic lesions.The five-stage agentic workflow uses these masks as its segmentation outputs.Source: §III · p. 2
- A computational layer derives volumes, lesion counts, bounding-box dimensions, and Couinaud locations from three-dimensional masks.Volumes use voxel counts and physical spacing; lesion localization uses the segment with the largest mask overlap.Source: §III · pp. 2–3
- The LLM organizes structured measurements into a fixed reporting schema, while schema validation and source-JSON checks verify structure and numerical consistency.Values without a computational source are flagged for review.Source: §III · p. 3
- The report is displayed with aligned segmentation overlays so radiologists can inspect findings and measurements, modify the report, and approve it.Reported measurements are linked to their computational sources for verification and review.Source: §III · pp. 2–3
IV. EXPERIMENTAL SETUP
The experimental setup evaluates segmentation and lesion detection on LiTS cases, compares lesion-level performance with published results, and includes a reporting prototype whose accuracy and usability were not quantitatively assessed.
IV. EXPERIMENTAL SETUP
- The LiTS dataset was used to evaluate liver and lesion segmentation on 51 CT cases, while MERLIN reports were screened exploratorily for liver abnormalities.Published Nicoli et al. results served as a reference, but cohort and protocol differences limit direct comparison.Source: §IV · p. 3
- Segmentation was assessed with the Dice similarity coefficient; lesion detection was assessed with sensitivity, positive predictive value, and false positives per case.The setup also includes a representative overlay comparing predicted and reference masks.Source: §IV · p. 3, Figure 3 · p. 3
- Reporting included numerical consistency checks against computed measurements, but reporting accuracy and clinical usability were not quantitatively evaluated.Source: §IV · p. 3
V. RESULTS
On LiTS, lesion detection achieved 75.6% sensitivity and 68.9% PPV, while liver Dice was 0.947 and lesion Dice was 0.63. MERLIN screening flagged 40.0% of reports as containing liver-related abnormalities, without independent labels to establish classification accuracy.
V-A. Segmentation Performance
- 0.947 liver Dice and 0.63 lesion Dice were achieved, with lesion segmentation reported as more challenging.All 51 evaluated LiTS cases passed orientation checks and were processed without flagged or failed cases.Source: §V · p. 3
V-B. Lesion Detection Performance
- The evaluated model achieved lesion sensitivity of 75.6%, PPV of 68.9%, and 0.70 false positives per case; published Nicoli et al. values were 65.7%, 80.17%, and 0.73, respectively.The results are not from a paired evaluation; differences are descriptive and do not represent a controlled comparison.Source: §V · p. 3, Table I · p. 4, Figure 4 · p. 4
V-C. Exploratory Report Screening
- 10,208 of 25,494 MERLIN reports (40.0%) were identified as containing liver-related abnormalities.These are predicted screening outcomes; classification accuracy could not be established without independently verified labels.Source: §V · p. 4
VI. DISCUSSION
The discussion presents reliable liver segmentation alongside more challenging lesion detection, and identifies larger matched-cohort and reporting evaluations as future work.
VI. DISCUSSION
- The LiTS findings indicate reliable liver segmentation, while lesion detection remains more challenging, reflected in the sensitivity–PPV trade-off.The discussion identifies the primary contribution as connecting segmentation outputs to deterministic lesion quantification and evidence-grounded structured reporting with numerical verification.Source: §VI · p. 4
- The current evaluation covers 51 LiTS cases, using published results as a reference rather than a direct comparison.Future work will use larger matched cohorts and assess measurement accuracy against reference annotations.Source: §VI · p. 4
- Future work will quantify reporting consistency and radiologist usability.These evaluations are proposed to assess the integrated workflow's potential for reproducible quantitative radiology reporting.Source: §VI · p. 4
VII. CONCLUSION
The conclusion presents a framework linking segmentation, quantitative analysis, and structured radiology reporting, with numerical checks and radiologist review. The system is a research prototype for supervised decision support, not autonomous diagnosis or clinical deployment.
VII. CONCLUSION
- The framework connects automated liver and lesion segmentation with quantitative analysis and structured radiology reporting.It separates image-derived measurements from LLM report composition and combines deterministic measurement tools, numerical consistency checks, and radiologist review.Source: §VII · p. 4
- The system is a research prototype for radiologist-supervised decision support, not autonomous diagnosis or clinical deployment.The study uses publicly available, de-identified LiTS and MERLIN datasets and involves no new patient data collection or re-identification.Source: ETHICS STATEMENT · p. 4