AI-based leukemia detection in routine diagnostic blood smear data

Introduction

Acute promyelocytic leukemia is an extremely aggressive blood cancer where immediate diagnosis can determine life or death. The diagnostic state of the art is manual inspection of a patient’s blood smear under the microscope by trained cytologists. It is prone to human error and time-consuming—a risk factor in notoriously understaffed laboratories. Supporting clinical decisions with AI will drastically increase diagnostic speed and accuracy, benefit patient survival, and free up valuable expert time.

Market Context

This is particularly important for cytological and histological analysis, whose market size is expected to rise by a compounded annual growth rate of 14.7% in the coming years. Yet, so far, the proof of concept that AI can be effectively employed for leukemia detection in routine diagnostics is missing.

Project Overview

I will leverage the methodological advancements in deep learning and explainable AI, the skills of my ERC CoG funded research group, and the expertise and data of the Munich Leukemia Laboratory (MLL), the largest leukemia laboratory in Europe and my longstanding industry partner.

Objectives

Together, we will develop and implement LeukoScreen, an AI-based software to automatically identify and flag up acute leukemia cases from MLL’s routine laboratory input. This will decrease the diagnosis to treatment time of critical leukemia cases at reduced costs and staffing. Specifically, we will:

1. Deploy a real-world dataset from the routine input of the MLL.
2. Train and evaluate our algorithm for transparent decision making on routine diagnostic blood smears.
3. Quantify the gain in sensitivity, specificity, and speed by comparing LeukoScreen with the currently used manual workflow at MLL.
4. Jointly develop a commercialization strategy for the exploitation of results.

Impact

This AI approach to support disease detection will save patients’ lives, change the paradigm of cytologic workflows, and create capacities in overburdened diagnostics.