Uveal Melanoma Cytology : Computer Assisted Diagnosis

Abstract

 

Purpose

 

To develop a fully automated, customized image processing algorithm to quantify the total number of cells in a given cytological slide and to detect spindle cells as an aid in the diagnosis of uveal melanoma.

 

Methods

 

Comparative masked study of ocular FNAB cytopathology slides prepared using standard laboratory techniques in a cytopathology laboratory. 10 cases of primary uveal melanoma and 5 cases of non-melanoma uveal tumors. The manual counts of all cells and spindle cells by an independent masked reader were recodered as standrad. The manual counts were compared to the automated counts (CAD) for each image tile to analyze the detection rates for any cell and spindle cells (Figure 1). The CAD algorithm was tuned to minimize the difference between the manual and computer assisted diagnosis (CAD) outputs for each image tile.

 

Results

 

The manual counts for any cell ranged from 5.7 to 401.8 / μm2 (mean = 112.8) for non-melanoma cases and 7.1 to 112.1 / μm2 (mean = 49.5) for melanoma cases. The computer assisted and the manual counts had an average cell detection difference of 9.9 ± 19.2 for non-melanoma cases and 4.0 ± 7.1 for the melanoma cases (average 9.8%). There was no significant correlation between the cellularity of the slide and the percentage difference in counts. The spindle cell detection difference between computer assisted and the manual counts was 5.2 ± 4.6 and -7.6 ± 6.8 for the non-melanoma cases and melanoma cases, respectively. The computer assisted counts calculated the average spindle cell composition rate for non-melanoma and melanoma cases to be 4.53 ± 2.43% and 23.04 ± 9.65%, respectively.

 

Conclusions

 

Automated digital methods can identify, quantify, and characterize spindle cells in an ocular FNAB cytopathology slide. Such automated system may be used to identify melanoma cases that could be verified by an expert cytopathologist at a remote site. Use of this computer assisted diagnostic tool can potentially be expanded for cytopathologic assessment of other tumors.

 

 

Figure 1. Manual count training of cell and spindle cells. (A) Large field-of-view image comprised of 41,412 um2 > field-of-view (15 grids) at 0.185 um/pixel resolution. (B) Magnified view of single grid with manual cell counts (cyan dots) and spindle counts (red dots). (C) Analysis output from CAD algorithm associated to manual counts tile in B with cell counts (green dots) and spindle counts (red outline).

 

Keywords: 589 melanoma • 639 pathology techniques • 624 oncology