Best imaging signs identified by radiomics could outperform the model: application to differentiating lung carcinoid tumors from atypical hamartomas

Insights into Imaging

Table 3 Confusion matrix for different thresholds on validation set with corrected prevalence

To predict hamartoma	True hamartoma	True carcinoid	Sensitivity	Specificity	cPPV	cNPV
3D < 10 HU			0.23 [0.10–0.43]	1.00 [0.89–1.00]	1.00 [0.24–1.00]	0.79 [0.74–0.83]
Predict hamartoma	5	0
Predict carcinoid	17	32
2D < 10 HU			0.13 [0.05–0.33]	1.00 [0.89–1.00]	1.00 [0.14–1.00]	0.77 [0.73–0.81]
Predict hamartoma	3	0
Predict carcinoid	19	32
Best threshold (To predict hamartoma)	True hamartoma	True carcinoid	Sensitivity	Specificity	PPV	NPV
3D < 40 HU			0.82 [0.61–0.93]	0.78 [0.61–0.89]	0.57 [0.35–0.75]	0.93 [0.82–0.97]
Predict hamartoma	18	7
Predict carcinoid	4	25
2D < 40 HU			0.82 [0.61–0.93]	0.81 [0.65–0.91]	0.60 [0.34–0.78]	0.93 [0.83–0.97]
Predict hamartoma	18	6
Predict carcinoid	4	26
To predict carcinoid	True carcinoid	True hamartoma	Sensitivity	Specificity	PPV	NPV
3D > 60 HU			0.63 [0.45–0.77]	0.95 [0.78–0.99]	0.82 [0.42–0.96]	0.88 [0.80–0.92]
Predict carcinoid	20	1
Predict hamartoma	12	21
2D > 60 HU			0.69 [0.51–0.82]	0.91 [0.72–0.98]	0.73 [0.39–0.94]	0.89 [0.81–0.94]
Predict carcinoid	22	2
Predict hamartoma	10	20

This table illustrates different confusion matrix for different thresholds (< 10 and > 60 HU) chosen on the training set, measured using 2D or 3D segmentations on the external validation set. Best threshold was chosen according to highest Likelihood Ratio (= 4.9). The corrected prevalence was set to 26%. HU—Hounsfield units; cNPV—corrected negative predictive values; cPPV—corrected positive predictive values