Towards Designing Accurate FISH Probe Detection using 3D U-Nets on Microscopic Blood Cell Images

Fluorescence in-situ hybridization (FISH) is a molecular cytogenetic technique developed to detect or localize the presence or absence of specific DNA sequences or chromosomes. Lung LB is a FISH based confirmatory diagnostic test for lung cancer which detects circulating tumor cells (CTC) in clinical patients with indeterminate lung nodules. In this paper, we propose a novel approach to segment FISH probes using 3D U-Nets and highlight the limitations of traditional Computer Vision based segmentation techniques for microscopic images. We observe a significant reduction in false positive rates without losing any real verified CTC, thus helping to improve the efficiency of the pathologists and accuracy of Lung LB. The proposed method results in a average percentage reduction of 62.875% in the number of falsely detected CTCs over the commercially available tool on 20 clinical cases (~1,86,901 cells), while achieving an average of 94.72% recall across the cases, showing an improvement over the the recall of 72.9% of the the commercial system.