Data-driven Cancer Research Conference 2026

Evaluating the Repeatability of MR-Derived Radiomic Features During Radiotherapy for Locally Advanced Cervical Cancer

Objectives:

Radiomics extract quantitative information from images

This information can be utilised to help predict treatment response or aid diagnosis

Prior to clinical implementation, radiomic features require assessment of repeatability

Temporal repeatability remains under researched [1]

This study aims to assess temporal repeatability of MR-derived radiomic features

Materials/Methods:

Materials:

Ten patients with locally advanced cervical cancer

Three visits, weeks 1, 3 and 5 of treatment

Two T2W MRI scans acquired per visit

Cervix-uterus and visible tumour delineated on each scan

Radiomic Feature Extraction:

Pre-processing was achieved via histogram matching normalisation

Radiomic Features were extracted using PyRadiomics (v3.0.1a)

An IBSI compliant open-source software

Repeatability Assessment:

Repeatability was assessed via intra-class correlation coefficient; ICC[2,1]

With scans as targets and contours as raters

Results:

Across all visits;

74.8% (cervix-uterus) 56.4%(visible tumour) of features showed excellent reliability (ICC > 0.9)

20.9% (cervix-uterus) 29.6%(visible tumour) good (0.75–0.9) 5.0% (cervix-uterus) 11.5% (visible tumour) moderate (0.5–0.75)

0.9% (cervix-uterus) 1.9% (visible tumour) features were labelled as poorly reliable

Systematic increases in repeatability across time were noted for the cervix-uterus, but not for the visible tumour, Table 1, Figure 1

Discussion/Future Work:

The majority (>90%) radiomic features are repeatable; denoted by an ICC > 0.5, Table 1

Features derived from the tumour less repeatable than those derived from the cervix-uterus

Radiomic feature repeatability changes across the treatment timeline, Figure 1

Longitudinal increase in cervix-uterus feature repeatability is theoretically driven by the shrinkage of the tumour

Future work for this project will establish an alternate feature selection pipeline which incorporates temporal information

Evaluating the Repeatability of MR-Derived Radiomic Features During Radiotherapy for Locally Advanced Cervical Cancer