# Prostate cancer tissue mapping and stratification using DRAQ5 and Eosin fluorescent labels integrated with AI classification and segmentation algorithms

## TL;DR

This paper presents a method using fluorescent labels and AI to classify and segment prostate cancer tissue, showing potential for clinical diagnostics.

## Contribution

The novel integration of DRAQ5 and Eosin fluorescent labeling with AI for prostate cancer tissue classification and segmentation.

## Key findings

- AI classifiers achieved high AUC scores for distinguishing healthy and cancerous prostate tissue.
- Segmentation models showed robustness against image acquisition variability.
- DRAQ5 and Eosin labeling combined with AI offers a promising pipeline for clinical prostate cancer diagnostics.

## Abstract

Fluorescent microscopy using the DRAQ5 and Eosin probes has been shown in the literature to be capable of producing rapid tissue characterization through synthetic H&E-like pseudoimages, which can be potentially utilized in the clinic. This study focuses on developing deep learning models for classification and segmentation of prostate tissue labeled with DRAQ5&Eosin. The fluorophores provide highly specific features of nuclear and cytoplasmic content that allows for enhanced spatial resolution and multi-parametric analytics. The inter-dependencies of image acquisition and configuration variability on AI predictive accuracy is systematically interrogated. We are thus able to establish limits on experimental and analytical robustness in automated Gleason Grading (1–5) tissue samples of prostate cancer.

A labeling technique based on a far-red DNA probe DRAQ5, and Eosin allowed us to generate a two-channel fluorescent readout of prostatic tissue samples. Deep learning networks were employed to classify and segment DRAQ5 and Eosin fluorescent image regions into healthy and high/low grade cancerous tissue. A subset of images were acquired with variable microscopy configurations (focus, noise, zoom, lens) to evaluate the robustness of the proposed experimental-analytical pipeline and reproducibility of predictions.

Machine Leaning classifiers of High Grade Cancer (Gleason pattern 4 or 5) vs Healthy, Low Grade Cancer (Gleason pattern 3) vs Healthy, and High Grade Cancer vs Low Grade Cancer achieved an area under the curve of 0.9314, 0.8398, and 0.7715 respectively. Pixel wide cancer segmentation attained DICE scores of 0.8436, 0.5138, and 0.705 for background, healthy, and cancerous tissue respectively. The segmentation model also displayed robustness against a broad range of induced acquisition variability.

Overall, DRAQ5 and Eosin labeling in combination with AI tools demonstrate a potential pipeline used in diagnostic clinical application when employing fluorescent imaging. Future research could expand and bring this combined fluorescent biomarker and AI methodology to the clinic.