Maschinelles Lernen zur Identifikation pathologischer Myokardregionen: ein innovativer Ansatz für die digitale Pathologie

Background/objectiveDigitalization and artificial intelligence (AI) are increasingly being tested and integrated into the evaluation of histological sections. While the application in the field of tumor pathology is already well advanced, experience with the digital analysis of histological sections of the myocardium is limited. We present a project for AI-supported analysis of histological samples in myocardial infarction, where we tested whether a specifically trained algorithm could distinguish healthy myocardial sections from those affected by ischemia. Material and methodsA total of 106 slides with HE-stained myocardial sections from 50 deceased individuals were digitalized and annotated. A convolutional neural network (CNN) based on a ResNet-18 architecture was trained for AI-assisted classification of the digitalized myocardial sections. The data were divided and slide-wise stratified into a training, validation and test dataset and analyzed by the deep learning algorithm for image pattern recognition of the CNN. ResultsThe algorithm developed was able to reliably distinguish healthy myocardium from pathological alterations in most cases during the test runs. Healthy myocardial tissue was detected with a precision of 81%, infarcted tissue with 78% and fibrosis with 85%. DiscussionThe insights gained in this project will be used to further develop the approach of AI-supported analysis of myocardial sections. With the expansion of the sample in terms of the number of cases and the pathologies present, future applications may arise in histological examinations of sudden cardiac death, especially in cases of underlying rare diseases (hereditary cardiomyopathies, etc.), where to date only limited micromorphological correlates could be identified using conventional histological analysis.