Metrics and Textural Features of MRI Diffusion to Improve Classification of Pediatric Posterior Fossa Tumors

Article Figures & Data

Download figure
Open in new tab
Download powerpoint
Fig 1.
T1WI+Gd (left), T2WI (middle), and ADC map (right) of an anaplastic (top) and classic MB. The overlaid region of interest (inside the green outline) is used to derive shape features and to calculate histogram and texture features for each image sequence.
Download figure
Open in new tab
Download powerpoint
Fig 2.
Training process to create single support vector machine classifiers for each tumor type and a combination step to produce a posterior fossa classifier to be tested on the remaining data.
Download figure
Open in new tab
Download powerpoint
Fig 3.
Distribution of 25th and 75th percentile values and skewness from ADC histograms for PAs, EPs, and MBs. White matter normalized values are used in the classifiers.
Download figure
Open in new tab
Download powerpoint
Fig 4.
Average classification rates for 2 PA tumor-type classifiers as a function of the number of bins used to quantize the normalized ADC values. Classification rates for histogram feature-based classifiers are independent of the number of bins (as long as these are not very low). In contrast, if the number of bins is too large, classification based on Gray-Level Co-Occurrence textural features will be affected (insufficient pairs of pixels are found).

Table 1:

Average tumor ADC values in pediatric posterior fossa tumors (×10⁻³mm²/s)

	Yamasaki et al 2005³	Rumboldt et al 2006⁶	Schneider et al 2007⁵	Bull et al 2012⁷
No. (MB/EP/PA)	(9/6/6)	(8/5/17)	(7/2/4)	(16/5/11)
EP	1.05–1.33	0.97–1.29	−0.8–1.4	1.10–1.25
MB	0.68–0.99	0.48–0.93	−0.5–1.0	0.67–1.22
Sig. difference (MB/EP)	Yes	Yes	No	No

Table 4:

Shape, histogram, and texture parameters used in analysis

Parameter
Shape	Volume, compactness, solidity
Histogram	Mean variance, mode, maximum probability, skewness, kurtosis, energy, entropy; percentiles: 10%, 25%, 50%, 75%, and 90%
Gray-Level Co-Occurrence Matrix	Autocorrelation, contrast, correlation, cluster prominence, cluster shade, dissimilarity, energy, homogeneity, maximum probability, sum of squares variance, sum average, sum variance, sum entropy, difference variance, difference entropy, information measure of correlation, inverse difference normalized, inverse difference moment normalized

Table 5:

Average tumor ADC values (×10⁻³mm²/s)

	PA	EP	MB	P Value^{^a}
Tumor
Mean/SD	1.70/0.26	1.34/0.29	0.85/0.18	<.05
Range	0.76–2.91	0.72–2.33	0.49–1.90
Normal-appearing white matter
Mean/SD	0.72/0.03	0.76/0.04	0.81/0.06	<.05
Range	0.63–0.87	0.62–0.95	0.59–0.93

↵a Between-group means comparison for the 3 groups (using 1-way ANOVA and Tamhane T2 post hoc multiple comparisons correction) were all significant.

Table 6:

Average correct classification rates for joint posterior fossa classifiers (3-PFT) based on shape, T2WI, and T1WI+Gd histogram/texture features^{^a}

T1WI+Gd and T2WI Features	PA (%)	MB (%)	EP (%)	3-PFT C (%)
ROI volume + T1WI+Gd histogram energy + T1WI+Gd sum entropy	83.5	78.2	74.6	78.8
ROI volume + T1WI+Gd mean + T1WI+Gd sum entropy	80.0	81.9	69.7	76.4
T2WI histogram skewness + T2WI mean + T2WI cluster prominence + T2WI sum variance	76.7	78.1	71.3	75.2

Note:—3-PFT indicates 3 posterior fossa tumor; 3-PFT C, 3 posterior fossa tumor classifiers.
↵a The performance of the separate individual classifiers that make up the combined classifier is also shown.

Table 7:

Average correct classification rates for joint posterior fossa classifiers (3-PFT) based on shape and ADC histogram/texture features^{^a}

ADC Features	PA (%)	MB (%)	EP (%)	3-PFT C (%)
Histogram 25th percentile + histogram 75th percentile + histogram skewness	96.9	95.8	94.3	91.4
Histogram 25th percentile + histogram median	95.6	92.0	91.8	89.2
ROI volume + histogram 75th percentile + histogram median + histogram entropy	96.2	91.3	83.9	87.4
Histogram 25th percentile	95.6	91.1	88.7	85.3
Histogram 75th percentile	96.1	89.7	85.1	83.5
Histogram median	92.3	89.9	82.2	78.9

Note:—3-PFT indicates 3 posterior fossa tumor; 3-PFT C, 3 posterior fossa tumor classifiers.
↵a The performance of the separate individual classifiers that make up the combined 3-PFT classifier is also shown. The bottom 3 rows correspond to the best single-feature classifiers.