Novel Deep Learning Pipeline for Automated Fluorescence in situ Hybridization Analysis

Abstract

Fluorescence in situ hybridization (FISH) is a laboratory technique used for detecting and locating a specific DNA sequence on a chromosome. It is one of the preliminary tests for diagnosis of many types of cancers and prenatal abnormalities. It is also used for identifying the course of treatment for cancer patients. A fluorochrome mix, also called a probe, is used to label parts of chromosomal regions or genes. These marked chromosomal regions appear as brightly colored spots (referred to as signals) against the DAPI stained blue nucleus. The presence (or absence), and counts of these spots in 100 to 200 cells translates to the result of the test. The analysis for the test is typically performed manually and requires two expert technologists for scoring and validation. The turnaround time for the test is very high and costs many hours in expert time. The technologist observes a FISH sample slide under the microscope and uses a manual counter to keep count of the signals and patterns visible in the nuclei. FISH slide quality deteriorates over time with the fluorescence intensity fading, and the same slide kept for over a month can not be analyzed again even if stored properly. There are several automated solutions available in the market like Abbott BioView, 3DHistech Pannoramic Scan, etc. The cost of these automated systems is in orders of millions of US dollars. Despite this, the automated analysis through these systems has been found to require considerable manual intervention. The proposed solution integrates a color camera with a traditional fluorescence microscope for capturing fields of view as the technologist observes the FISH slide under the microscope. The captured images are analyzed by a robust deep learning pipeline to deliver the final scores of the patterns observed in the nuclei. The raw images are captured at multiple focus levels, preprocessed and stacked together to form a final composite image for each field of view. The stacked images then pass through a multiclass nucleus detector that detects the viable and non viable nuclei of different types. The viable nuclei are fed to a segmenter that extracts the signal boundaries. A signal classifier then assigns a color label to each of these signals. For the sake of ease of analysis, each nucleus is transformed into a pseudo or false image that has a very bright nucleus background 5 with clearly visible signals. The count and type of signal observed against every nucleus is then reported as the final score. The type of abnormal patterns for sample and probe types are known. Laboratory cut-offs for each probe pattern are used to classify observed patterns as abnormal or normal. The final output of the pipeline is pattern scores in percentages and pseudo images, which delivers these results in order of milliseconds, reducing the analysis time from hours to seconds.