A Combined Radiomics and Machine Learning Approach to Overcome the Clinicoradiologic Paradox in Multiple Sclerosis

Article Figures & Data

Download figure
Open in new tab
Download powerpoint
FIG 1.
Flow diagram showing inclusion and exclusion criteria.
Download figure
Open in new tab
Download powerpoint
FIG 2.
Workflow summarizing the main MR imaging data-processing and data-mining steps. Image illustrates the data set composition as well as the major steps performed for feature extraction, feature selection, and regression modeling. LL indicates lesion load; NAWM, normal-appearing white matter; ChaCo, change in connectivity; T1w, T1-weighted imaging.

Table 1:

Clinicodemographic characteristics of the studied population, along with MR imaging–derived global brain volumes^{^a}

	Site 1 (n = 500)	Site 2 (n = 104)	P Value (Site 1 vs Site 2)
Age (mean) (yr)	37.5 (SD, 10.9)	38.3 (SD, 9.8)	.49
Female sex (No.) (%)	349 (69.8)	80 (76.9)	.16
Secondary-progressive course (No.) (%)	72 (14.4)	20 (19.2)	.23
DD (mean) (yr)	9.3 (SD, 8.1)	9.2 (SD, 8.5)	.83
EDSS (median) (IQR)	2.5 (2.0–4.0)	2.0 (1.5–4.0)	.03
TLV (mean) (mL)	10.6 (SD, 13.4)	7.2 (SD, 8.6)	.05
WBV (mean) (mL)	1026.1 (SD, 116.3)	1042.6 (SD, 117.4)	.48

Note:—TLV indicates total lesion volume.
↵a Between-group differences regarding MR imaging measures are adjusted for age, sex, and estimated total intracranial volume.

Table 2:

Selected radiomics features^{^a}

Anatomic Label	Feature Class	Class Characteristics	Feature	Feature Characteristics
Right frontal superior orbital cortex	First order	Describes the distribution of voxel intensities	Median	The median gray level intensity
Left amygdala	Gray level co-occurrence matrix	Quantifies how often pairs of pixels with specific values occur in a specified spatial range	Correlation	Measures the linear dependency of gray level values to their respective voxels in the matrix
Left caudate nucleus	Gray level co-occurrence matrix	Quantifies how often pairs of pixels with specific values occur in a specified spatial range	Informational measure of correlation 1	Quantifies the complexity of the texture
Right thalamus	First order	Describes the distribution of voxel intensities	Energy	Measures the magnitude of voxel values
Left cerebellar lobule VIII	Gray level dependence matrix	Quantifies gray level dependencies (ie, the number of connected voxels within a set distance that are dependent on the center voxel)	Small dependence low gray level emphasis	Measures the joint distribution of small dependence with higher gray-level values
Cerebellar vermis (lobules IV–V)	Gray level size-zone matrix	Quantifies gray level zones (ie, the number of connected voxels sharing the same intensity value)	Size-zone non-uniformity	Measures the variability of size-zone volumes
Left cerebellar crus	First order	Describes the distribution of voxel intensities	Median	Median gray level intensity

↵a Characteristics of each selected feature and relative class according to PyRadiomics official documentation (https://pyradiomics.readthedocs.io/en/latest/features.html) are presented, along with the anatomic location (according to Tzourio-Mazoyer et al²⁰) of the corresponding ROI.

Table 3:

Machine learning predictive models^{^a}

Cohort	Ridge Regression			Gaussian Process			Support-Vector Machine			Random Forest			P Value
Cohort	r	R²	MAE	r	R²	MAE	r	R²	MAE	r	R²	MAE	P Value
Training	0.797	0.636	0.651	0.795	0.632	0.814	0.797	0.635	0.710	0.790	0.624	0.656	–
Internal test	0.741	0.549	0.725	0.733	0.537	0.874	0.734	0.538	0.754	0.734	0.539	0.740	.001^{^a}
External test	0.755	0.570	1.155	0.799	0.638	1.247	0.794	0.631	1.112	0.775	0.600	1.162	.009^{^b}

Note:—– indicates not available.
↵a F(1.82, 180.39) = 7.94. Partial η² = 0.07. df corrected using Greenhouse-Geisser estimates of sphericity (ε = 0.61).
↵b F(1.70, 175.52) = 5.25. Partial η² = 0.05. df corrected using Greenhouse-Geisser estimates of sphericity (ε = 0.57).