Intracerebral Hemorrhage Detection in Computed Tomography Scans Through Cost-Sensitive Machine Learning

Figures & data

Figure 1. Sample dataset images.

Table 1. Parameter specifications.

Name	Parameters
Random Forest	FeatureFraction = $\frac{1}{4\sqrt{10}}$, LeafSize = 5, TreeNumber = 50, DistributionSmoothing = 0.5
Decision Tree	DistributionSmoothing = 1, FeatureFraction = 1
Gradient Boosted Trees	BoostingMethod = Gradient, MaxTrainingRounds = 50, LeavesNumber = 13, LearningRate = 0.1, MaxDepth = 6, LeafSize = 15, L1Regularization = 0, L2Regularization = 0
Nearest Neighbors	NeighborsNumber = 5, DistributionSmoothing = 0.5, NearestMethod = Scan
Support Vector Machine	KernelType = RadialBasisFunction, GammaScalingParameter = 0.00113124, PolynomialDegree = 3, SoftMarginParameter = 1, BiasParameter = 1, MulticlassStrategy = OneVersusOne
Logistic Regression	L1Regularization = 0, L2Regularization = 100, OptimizationMethod = LBFGS

Table 2. Utility matrix used for this study.

	Hemorrhage	No Hemorrhage
Hemorrhage	$1$	-$x$
No Hemorrhage	$0$	$1$

Figure 2. Probability histograms.

Figure 3. Trends for sensitivities and specificities by penalty.

Table 3. Round 1 results.

	Accuracy	Sensitivity	Specificity
Decision Tree	0.600	0.800	0.400
Random Forest	0.700	0.750	0.650
Gradient Boosted Trees	0.900	0.900	0.900
Nearest Neighbors	0.825	0.700	0.950
Support Vector Machine	0.975	0.950	1.00
Logistic Regression	0.925	0.900	0.950

Table 4. Round 2 results.

	Penalty of −1			Penalty of −2			Penalty of −3
	Acc	Sens	Spec	Acc	Sens	Spec	Acc	Sens	Spec
Decision Tree	0.500	1.00	0	0.500	1.00	0	0.500	1.00	0
Random Forest	0.650	0.900	0.400	0.650	0.900	0.400	0.700	1.00	0.400
Gradient Boosted Trees	0.900	1.00	0.800	0.900	1.00	0.800	0.850	1.00	0.700
Nearest Neighbors	0.800	0.900	0.700	0.800	0.900	0.700	0.625	0.950	0.300
Support Vector Machine	0.925	1.00	0.850	0.925	1.00	0.850	0.900	1.00	0.800
Logistic Regression	0.925	0.950	0.900	0.900	0.950	0.850	0.850	1.00	0.700

Data Availability Statement

The dataset employed was provided under a CC0: Public Domain license by Kitamura (2017). The code for the study can be accessed at https://github.com/rushankgoyal/cost-sensitive-hemorrhage.