Automated clinical decision support for determining HER2 gene amplification from FISH images in breast cancer

Fluorescence in situ hybridization is required in cases of equivocal immunohistochemical assessment. Visual analysis of such images is labor-intensive and subjective due to specimen heterogeneity, as well as technical and biological artifacts. To ensure reproducibility of results and reduce the physician’s workload, there is a need for an interpretable automated algorithm based on guidelines for manual analysis. The goal of this study is to investigate the specific features of applying manual analysis guidelines in automated clinical decision support using binary segmentation of nuclei and signals in fluorescence in situ hybridization images, and to assess the influence of additional features of the segmented regions on algorithm performance. In a dataset of 155 images from 44 patients, nuclei and signal segmentation was performed using pretrained deep learning models. It was shown that the basic adaptation of manual analysis guidelines provides agreement with expert assessment at the level of κ = 0.36. The use of a weighted signal ratio and optimization of the threshold value increased agreement to κ = 0.49. The use of a meta-classifier based on the number, area, and intensity features of the segmented objects further increased agreement to κ = 0.58. The proposed approach makes it possible to improve the objectivity and reproducibility of fluorescence in situ hybridization image interpretation in breast cancer and to reduce the labor intensity of the analysis for the physician.