European Radiology: Prediction of preoperative T stage of rectal cancer by a 3D super-resolution MRI radiomics model based on deep learning

Currently, colorectal cancer ranks third and second
in the global spectrum of tumor incidence and death, respectively.
Rectal cancer (RC) accounts for 1/3 of colorectal cancers and has the highest
incidence in East Asia.
Although radical surgery remains the only cure, the implementation of neoadjuvant chemotherapy (nCRT) based on pre-treatment staging has increased the rate of R0 resection and further improved local tumor control
.
nCRT is recommended as standard of care for patients with locally advanced RC (LARC) (T3/4 and/or N+), while there is no indication for treatment in patients with early stage (T1/2 and N-
).
To avoid overtreatment or undertreatment, accurate preoperative assessment of the T stage to distinguish between T1/2 and T3/4 tumors is essential
.
High-resolution (HR) T2-weighted imaging (T2WI) has fine soft tissue resolution and has become the first-line imaging method of RC office in partial stages.
However, overall accuracy ranged from 60 to 75%, with studies reporting an average overstaging rate of 30-57%
for T2 tumors.
In this case, improving the evaluation of the T stage of the primary tumor has important clinical value
for precise treatment and optimization of treatment strategies.

As an emerging data extraction method, radiomics can quantitatively assess tumor heterogeneity
by mining high-dimensional data in medical images.
Studies have shown that radiomics can improve patient management and clinical decision-making
by discovering disease features invisible to the human eye.
In RC, radiomics has performed impressively in different oncology scenarios, including assessing the biological behavior of tumors, assessing response to treatment, and predicting prognosis
.
Despite these advances, radiomic features are often influenced
by anisotropic resolution and low voxel statistics in current medical imaging.
In order to improve the robustness and stability of radiomics models, it is quite important
to address these limitations by applying higher resolution images in model construction.

Super-resolution (SR) technology, which aims to recover higher spatial resolution of digital images from low-resolution observations, has been proposed and introduced into the clinic since the 80s of the
20th century.
In recent years, with the development of deep learning (DL), SR has achieved excellent performance
in medical imaging.
In addition to the recovered SR images showing good stability and reliability in multispatial trapezoids, DL-SR has also attracted attention in medical imaging, as radiomic features extracted from SR images have been quantitatively demonstrated to have obvious repeatability and stability
.
However, none of the current radiomics studies have utilized DL-SR technology to discover specific radiogramic biomarkers for clinical use
.

Recently, a study published in the journal European Radiology developed and validated an SR radiomics model for preoperative prediction of T stage in RC patients, the first DL-based three-dimensional (3D) SR radiomics study
to be demonstrated in the clinic.

A total of 76 eligible RC patients (T1/2=287, T3/4=419) were included and allocated chronologically as training cohorts (n=565) and validation cohorts (n=141).

In this study, a deep transfer learning network analysis was performed on high-resolution (HR) T2-weighted imaging (T2WI) to improve the Z-resolution of images, and preoperative SRT2WI was obtained
.
The radiomics models named modelHRT2 and modelSRT2 were constructed
from high-dimensional quantitative features extracted by the minimum absolute contraction and selection combinator methods in artificially segmented volumes of interest of HRT2WI and SRT2WI, respectively.
The performance of the model is evaluated
by ROC, calibration, and decision curves.

ModelSRT2 was superior to modelHRT2 in differentiating T1/2 from T3/4 RC (AUC 0.
869, sensitivity 71.
1%, specificity 93.
1%, accuracy 83.
3% and AUC 0.
810, sensitivity 89.
5%, specificity 70.
1%, accuracy 77.
3%), the difference was significant (P < 0.
05).

Both radiological models achieved higher AUCs than radiologists (0.
685, 95% confidence interval 0.
595-0.
775, P < 0.
05).

The calibration curve confirmed the high fit, and the decision curve analysis confirmed its clinical value
.

Graph model evaluation
.
Cumulative distribution curve
of A Rad scores.
ROCs
for training (B) and testing (C) cohorts.
The calibration curve shows a good calibration of the model in both cohorts (D-G)

This study proposes a DL-based three-dimensional SR radiomics model that outperforms traditional HRT2 radiomics models and radiologists
in predicting tumor T staging and identifying nCRT candidates in RC patients.
This is the first time so far that DL-based 3D SR methods have been applied to radiomic analysis
for clinical decision support.
With further validation, the clinical utility of radiomics through SR imaging will be maximized
.

Original source:

Min Hou,Long Zhou,Jihong Sun.
Deep-learning-based 3D super-resolution MRI radiomics model: superior predictive performance in preoperative T-staging of rectal cancer.
DOI:10.
1007/s00330-022-08952-8