Predicting response to somatostatin analogues in acromegaly: machine learning-based high-dimensional quantitative...

To investigate the value of machine learning (ML)-based high-dimensional quantitative texture analysis (qTA) on T2-weighted magnetic resonance imaging (MRI) in predicting response to somatostatin analogues (SA) in acromegaly patients with growth hormone (GH)-secreting pituitary macroadenoma, and to compare the qTA with quantitative and qualitative T2-weighted relative signal intensity (rSI) and immunohistochemical evaluation.

Forty-seven patients (24 responsive; 23 resistant patients to SA) were eligible for this retrospective study. Coronal T2-weighted images were used for qTA and rSI evaluation. The immunohistochemical evaluation was based on the granulation pattern of the adenomas. Dimension reduction was carried out by reproducibility analysis and wrapper-based algorithm. ML classifiers were k-nearest neighbours (k-NN) and C4.5 algorithm. The reference standard was the biochemical response status. Predictive performance of qTA was compared with those of the quantitative and qualitative rSI and immunohistochemical evaluation.

Five hundred thirty-five out of 828 texture features had excellent reproducibility. For the qTA, k-NN correctly classified 85.1% of the macroadenomas regarding response to SAs with an area under the receiver operating characteristic curve (AUC-ROC) of 0.847. The accuracy and AUC-ROC ranges of the other methods were 57.4-70.2% and 0.575-0.704, respectively. Differences in predictive performance between qTA-based classification and the other methods were significant (p < 0.05).

The ML-based qTA of T2-weighted MRI is a potential non-invasive tool in predicting response to SAs in patients with acromegaly and GH-secreting pituitary macroadenoma. The method performed better than the qualitative and quantitative rSI and immunohistochemical evaluation.

• Machine learning-based texture analysis of T2-weighted MRI can correctly classify response to somatostatin analogues in more than four fifths of the patients. • Machine learning-based texture analysis performs better than qualitative and quantitative evaluation of relative T2 signal intensity and immunohistochemical evaluation. • About one third of the texture features may not be excellently reproducible, indicating that a reliability analysis is necessary before model development.