| Problem under investigation. Human-based semi-automated systems have remained the "gold standard" for machine-vision systems in measuring nuclear characteristics though there have been recent increases in the use of automated imaging systems. Both human-based and automated systems deserve further research into their weaknesses and strengths. The effect of excess variability associated with the human component in these systems has received little attention. Also, there has been little evaluation of directly comparable results from both human-based and automated systems.;Objective. First, the variability that was associated with the human component of a semi-automated machine-vision system was measured and assessed. Second, the ability to discriminate benign, pre-cancerous, and cancerous prostate tissues using measurements of morphological nuclear characteristics was compared using semi-automated and automated machine-vision systems.;Study design. A hematoxylin-eosin (H&E) stained section of prostate tissue microarray was used to construct image sets of sampled nuclei. Segmentation of these images was performed by trained human segmenters for use in the semi-automated system; segmentation in the automated system was an integrated process. Measured morphometric nuclear features included mean optical density (MOD), nuclear area (NA) and nuclear roundness factor (NRF).;Results. Overall differences between 4 segmenters using the semi-automated system was mixed, with NRF attaining significance (p<.001). NA was significant among sessions (p<.0009) and these segmenter differences also varied significantly (p<.0001). Similarly, differences in MOD varied among sessions (p<.0001) as well as within sessions (p<.049), though the magnitude of differences was small. NRF displayed differences among sessions (p<.0001), variance in these differences (p<.0001), and variance in intra-session differences (P=.026). To compare the semi-automated and automated systems, nuclei from 17-BPH, 4-HGPIN and 8-CaP glands were analyzed. Using multivariate models, both systems had comparable predictive abilities: the manual system displayed a greater ability to distinguish BPH and HGPIN (p<0.0001), while the automated system better differentiated BPH and CaP (p=0.01). In univariate models, the manual system distinguished better BPH and HGPIN using NA (p<0.0001) and MOD (p<0.0001), while the automated system distinguished better BPH and CaP using MOD (p<0.0001) and NRF (p=0.004).;Conclusions. Results suggest that human-based segmentation systems may be prone to statistically significant variation within morphometric measures. Since this variation is undetectable unless the investigator specifically employs means for detecting it, current practices and quality assurance methods should be reviewed. Additionally, the study results show that even though the automated system was developed for a highly specific nuclear stain application, it performed at least as well as human based systems with a more commonly used stain (H&E). The more limited human effort required for automated image analysis, alongside greater reproducibility, combine to make automated systems superior to manual systems for this application. |
