Repeatability of right ventricular segmentation among serial exams between post-processing methods

Background: Accurate and reproducible segmentation of ventricular volumes during post-processing of CMR studies is crucial for diagnosis and management of cardiac diseases. Although many software packages have optimized automatic segmentation of the left ventricle (LV), right ventricular (RV) segmentation remains challenging due to its unique geometry.¹ This is especially important for serial examinations where differences could lead to prognostic or treatment changes. We sought to investigate the impact of inter-observer repeatability of RV segmentation on serial CMR evaluation using a combination of post-processing techniques.

Methods: We employed 3 methods of post-processing using commercial software (SuiteHEART, Neosoft Inc.): (1) entirely automated analysis, (2) semi-automated analysis with expert corrections of automated segmentation (similar to clinical practice), and (3) entirely manual analysis by 3 experts. We calculated inter-observer correlations (ICC) for repeatability² and the minimum detectable difference (MDD)³ in RV volumes determined by the test/retest variability from 16 normal subjects. One set of cine images were acquired followed by a second set with a 10 minute break, where the normal volunteer was removed from the MRI. We applied this method to a prospectively recruited group of 24 oncology & 95 cardiomyopathy patients who had serial CMR examinations to determine the impact of variability from post-processing methods to detectable clinical changes of RV volumes.

Results: In the test/retest study from 16 normal volunteers, manual post-processing had the best repeatability and the smallest MDD attributed to segmentation variability (Table 1). Using RVEDV for example, a serial change of < 22 mL measured manually, < 29 semi-automatically, or < 32 mL automatically would be attributed to variability due to post-processing technique alone. MDD for RVESV was much less varying (MDD of 13 mL, 15 mL, and 16 mL for semi-automated, automated, and manual methods, respectively). RV stroke volume and ejection fraction (EF) demonstrated a similarly large distribution of variability as EDV. Applying MDD to serial cardio-oncology and cardiomyopathy cohorts (Table 2), we found a decreasing frequency of detectable changes using automated segmentation (23%) followed by semi-automated (32%) and manual methods (42%) in the case of RVEDV for cardiomyopathy patients. Similar patterns were observed for RV stroke volume and RVEF. Among oncology patients the frequency of interval detectable changes of RVEDV and RVEF was also at the lowest with automated, followed by semi-automated and manual segmentation.

Conclusion: Repeatability analyses of test/retest studies from normal volunteers found that manual segmentation of the RV resulted in smaller minimum detectable changes allowing clinicians to detect smaller changes between serial exams among patients with cardiomyopathy or at risk for chemotoxicity.

Table 1 Repeatability statistics of right ventricular segmentation in test-retest studies among normal volunteers

RV Measurement	Method	Mean 1	Mean 2	Mean difference	ICC	MDD
End diastolic volume	Automatic	148.6 ± 37.8	150.9 ± 40.6	-2.3 ± 16	0.92 (0.79, 0.97)	31.4
	Auto + adjustment	146.2 ± 39.5	150 ± 37.7	-3.8 ± 13.1	0.92 (0.8, 0.97)	29.3
	Manual	149.4 ± 45.9	149.8 ± 40.8	-0.3 ± 11	0.97 (0.92, 0.99)	21.6
End systolic volume	Automatic	65.9 ± 25.2	71.9 ± 27.4	-6 ± 7.8	0.94 (0.71, 0.98)	15.2
	Auto + adjustment	60.3 ± 22.3	62.6 ± 23.2	-2.3 ± 6.2	0.95 (0.86, 0.98)	12.8
	Manual	59.9 ± 23.5	62.1 ± 22.6	-2.2 ± 8	0.94 (0.84, 0.98)	15.8
Stroke volume	Automatic	82.7 ± 18.8	79 ± 17.6	3.7 ± 15.4	0.64 (0.25, 0.86)	30.1
	Auto + adjustment	85.9 ± 19.6	87.4 ± 16.1	-1.5 ± 9.8	0.82 (0.57, 0.93)	23
	Manual	88.4 ± 24.8	87.6 ± 20.3	0.8 ± 6.9	0.96 (0.88, 0.98)	13.5
Ejection fraction	Automatic	56.4 ± 7.7	53.2 ± 6.8	3.3 ± 6.1	0.60 (0.18, 0.84)	11.9
	Auto + adjustment	59.5 ± 5.1	59.2 ± 5	0.3 ± 3	0.79 (0.49, 0.92)	6.7
	Manual	59.7 ± 5.3	59.2 ± 4.8	0.5 ± 1.8	0.94 (0.83, 0.98)	3.5

 ICC = inter-class correlation coefficient; MDD = minimum detectable differenceTable 2. Detectable changes in RV volumes in serial studies among cardiomyopathy and cardio-oncology cohorts

RV Measurement	Change between serial studies			Detectable changes by post-processing method
	Mean	Min	Max	Automatic	Auto + adjustment	Manual
Cardiomyopathy (N=95)
End diastolic volume	1.1 ± 28	-51	83	22 (23%)	30 (32%)	43 (45%)
End systolic volume	1.3 ± 14.6	-27.2	55.3	21 (22%)	31 (33%)	19 (20%)
Stroke volume	-0.3 ± 19.7	-50.5	60.4	13 (14%)	25 (26%)	42 (44%)
Ejection fraction	-0.5 ± 6.7	-24	14	6 (6%)	29 (31%)	58 (61%)
Cardio-oncology group (N=24)
End diastolic volume	7.6 ± 27.9	-51	83	7 (29%)	8 (33%)	12 (50%)
End systolic volume	3.1 ± 11.9	-27.2	55.3	2 (8%)	7 (29%)	2 (8%)
Stroke volume	4.6 ± 22.8	-50.5	60.4	5 (21%)	5 (21%)	11 (46%)
Ejection fraction	-0.5 ± 6.9	-24	14	3 (13%)	8 (33%)	14 (58%)