Computerized decision support system for intraoperative analysis of margin status in breast conservation therapy

Abstract

Background. Breast conservation therapy (BCT) is the standard treatment for breast cancer; however, 32-63% of procedures have a positive margin leading to secondary procedures. The standard of care to evaluate surgical margins is based on permanent section. Imprint cytology (IC) has been used to evaluate surgical samples but is limited by excessive cauterization thus requiring experienced cytopathologist for interpretation. An automated image screening process has been developed to detect cancerous cells from IC on cauterized margins. Methods. IC was prospectively performed on margins during lumpectomy operations for breast cancer in addition to permanent section on 127 patients. An 8-slide training subset and 8-slide testing subset were culled. H&E IC automated analysis, based on linear discriminant analysis, was compared to manual pathologist interpretation. Results. The most important descriptors, from highest to lowest performance, are nucleus color (23%), cytoplasm color (15%), shape (12%), grey intensity (9%), and local area (5%). There was 100% agreement between automated and manual interpretation of IC slides. Conclusion. Although limited by IC sampling variability, an automated system for accurate IC cancer cell identification system is demonstrated, with high correlation to manual analysis, even in the face of cauterization effects which supplement permanent section analysis.

Figure 1

Representative imprint cytology specimens. (a) Clustered cancer cells with associated red blood cells. (b) Noncancerous individual white blood cells. (c) Debris/artifacts. Cancer cells are distinguished from other objects by their round shape, large size, staining color of the nucleus and cytoplasm, and their grouping in clusters.

Figure 2

Class separation in LDA-transformed space showing cancer (red), Noncancer cells (Green), and debris/artifact objects (Blue) clustered separately. 80% normal probability ellipses for each class are overlaid. Relative contributions by class of descriptor to class separation are listed. Nuclear color provides the best discrimination.

Figure 3

Top 15 (from 25) automatically identified suspicious cancer areas from sample T2. Cancer cells are present on areas 1, 2, 3, 4, 6, and 7. Additional misidentified noncancer objects were selected: imaging artifact at edge of slide (area 5), cautery debris (8, 9, 14), adipose (12, 14), and air bubble artifacts (8, 11, 12, 14, 15). Note: blue markings on area 7 are from manual evaluation and had no effect on automated slide evaluation.

Figure 4

Identification of cancerous clusters across representative imprint cytology specimen T2. Cancer area 2 was determined by both automated and manual imprint cytology. Two additional foci of cancer (Cancer area 1 and 3) were additionally identified by automated imprint cytology evaluation.