European Radiology: The value of radiomics models in predicting temporal lobe radiation damage in nasopharyngeal carcinoma

Nasopharyngeal carcinoma (NPC) is a very common endemic tumor in Southeast and East Asia, and radiation therapy is currently the preferred treatment for non-metastatic nasopharyngeal carcinoma, but the temporal lobe (TL) is inevitably included in the
radiation field.
The incidence of radiation-induced temporal lobe injury (RTLI) in patients receiving radiation therapy (IMRT) has been reported to be 4.
6% to 8.
5%.

Patients who develop RTLI experience impaired memory, language, and mobility, but nearly half of patients (45.
3%) have no symptoms of RTLI at diagnosis, and most have no symptoms
even in advanced stages.

At present, the imaging diagnosis of RTLI mainly relies on MRI.

However, existing conventional MRI techniques can only distinguish RTLI
at an irreversible stage.
Recently, artificial intelligence (AI) such as radiomics has been widely used to predict treatment outcomes
such as complications and disease progression.
Radiomics describes the process of extracting a large number of image-based features from routine diagnostic scans, and high-dimensional data that quantifies tumor shape, image intensity, and texture can reflect the characteristics of the disease, which is widely
used in clinical decision-making.

Recently, a study published in the journal European Radiology developed and validated a radiomics-based pre-treatment prediction model for predicting RTLI in NPC patients, providing technical support
for early clinical identification of the extent and risk of RTLI.

A total of 216 patients diagnosed with nasopharyngeal carcinoma were reviewed
.
All patients were randomly assigned to training (n = 136) and validation groups (n = 80).

Radiomic features
were extracted from pre-treatment T1-enhanced or fat-inhibiting T2-weighted MRI.
Radiomic features were generated by the Minimum Absolute Contraction and Selection Operator (LASSO) regression algorithm, Pearson correlation analysis, and univariate logic analysis, and clinical features
were selected using logistic regression analysis.
Multivariate logistic regression analysis was performed to establish three models for RTLI prediction of the training cohort: radiomic features, clinical variables, and clinical-radiomic parameters
.
A radiomic nomogram was used and calibrated, identified, reclassified, and evaluated for clinical application
.

This study found that a radiomic signature model consisting of two radiomic features was significantly associated
with RTLI.
The proposed radiomics model showed good discriminating ability in both training (AUC, 0.
89) and validation cohort (AUC, 0.
92), and exceeded the clinical prediction model (P < 0.
05).

Models combined with radiomics and clinical features have achieved higher AUC (AUC, 0.
93 and 0.
95) as well as better calibration and improved prediction accuracy
of RTLI.
Clinical-radiomic models have also shown excellent performance
in predicting RTLI in different clinical-pathological subgroups.

Figure (a) constructed a radiomic nomogram in the training cohort and incorporated radiomic score, T stage, and age
.
(b) Training cohort and (c) Validation cohort radiology nomogram calibration curves in the cohort

In this study, a machine learning method was developed and validated to predict and evaluate
radiation-induced temporal lobe injury (RTLI) in patients with nasopharyngeal carcinoma (NPC) through pre-treatment temporal lobe MRI images.
The identified radiomic profile can be used as a biomarker for RTLI risk stratification
.

Original source:

Dan Bao,Yanfeng Zhao,Lin Li,et al.
A MRI-based radiomics model predicting radiation-induced temporal lobe injury in nasopharyngeal carcinoma.
DOI:10.
1007/s00330-022-08853-w