Oops, you're using an old version of your browser so some of the features on this page may not be displaying properly.

MINIMAL Requirements: Google Chrome 24+Mozilla Firefox 20+Internet Explorer 11Opera 15–18Apple Safari 7SeaMonkey 2.15-2.23

Poster Display session

30P - AI-based response monitoring in gastroenteropancreatic neuroendocrine tumors

Date

15 Mar 2024

Session

Poster Display session

Presenters

Kalina Chupetlovska

Citation

Annals of Oncology (2024) 9 (suppl_2): 1-5. 10.1016/esmoop/esmoop102414

Authors

K. Chupetlovska1, J. Engel2, R. Beets-Tan1, M. Tesselaar3, M. Maas1, S. Trebeschi1

Author affiliations

  • 1 Radiology Dept., NKI-AVL - Netherlands Cancer Institute/Antoni van Leeuwenhoek Hospital, 1066 CX - Amsterdam/NL
  • 2 Technical Medicine, University of Twente, 7522 NB - Enschede/NL
  • 3 Medical Oncology Department, NKI-AVL - Netherlands Cancer Institute/Antoni van Leeuwenhoek Hospital, 1066 CX - Amsterdam/NL

Resources

This content is available to ESMO members and event participants.

Abstract 30P

Background

Neuroendocrine tumors (NETs) display relatively slow growth and heterogeneous treatment responses. Recent advances in systemic and locoregional therapies have shown efficacy in delaying progression. Imaging, while crucial for follow-up, faces challenges with conventional evaluation criteria (RECIST), criticized for variability in target lesion selection and measurements. AI-enabled automated segmentations may provide a time-efficient, accurate, and reproducible method for treatment response monitoring.

Methods

Two AI segmentation models were trained to delineate liver metastases in NET patients, following an iterative training-radiologist correction scheme: a portal venous phase (PVP) model (trained on 65 scans) and an arterial phase (AP) model (trained on 36 scans). AP model was trained in three different settings. The training dataset was from a previously published study. External validation involved testing the final models on an independent dataset with follow-up data from three exemplary patients with long follow-ups. The evaluation utilized the Dice Correlation Coefficient Score (DSC) quantifying the overlap between predicted and ground truth, coefficient of determination (r2).

Results

PVP model DSCs were 66%, 77%, and 93% for each patient, respectively. Correlation between AI volume and radiologist volume was high per lesion (r2=0.93), and total tumor burden (r2=0.79). The three settings used for the AP models showed similar results in terms of median DSC (87%, 88%, 91%). Both PVP and AP models exhibited variability in DSC due to lesion density and enhancement differences and heterogeneity, with PVP excelling in identifying the hypodense lesions and AP performing better in hyperdense lesions. Differences were observed between RECIST and AI-volume trends in a preliminary comparison.

Conclusions

This pilot study shows promising preliminary results for the creation of response criteria based on AI-quantified volumetry in NET tumors, with the algorithm being able to reproduce the results of experienced radiologists. Future outlook includes extension of the study cohort and formal comparison with RECIST 1.1.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.

This site uses cookies. Some of these cookies are essential, while others help us improve your experience by providing insights into how the site is being used.

For more detailed information on the cookies we use, please check our Privacy Policy.

Customise settings
  • Necessary cookies enable core functionality. The website cannot function properly without these cookies, and you can only disable them by changing your browser preferences.