Translational Medicine I. (Poster discussion will take place on the terrace of the room during the Coffee Break)
Csenger Kovácsházi1, András Horváth2, Csanád Tabajdi2, Adrienn Rácz2, Gábor B. Brenner1, Tamás G. Gergely1, Bence Ágg1,3,4, Péter Ferdinandy1,3,4, Anikó Görbe1,3,4, Zoltán Giricz1,3
1. Department of Pharmacology and Pharmacotherapy, Semmelweis University, 1089 Budapest, Hungary
2. Faculty of Information Technology and Bionics, Pázmány Péter Catholic University, 1083 Budapest, Hungary
3. Pharmahungary Group, 6722 Szeged, Hungary
4. MTA-SE System Pharmacology Research Group, Department of Pharmacology and Pharmacotherapy, Semmelweis University, 1089 Budapest, Hungary
Introduction: Planimetric measurement of myocardial infarct size (IS) by Evans blue and 2,3,5-triphenyltetrazolium (TTC) staining is widely used in preclinical studies. However, manual analysis of the images takes considerable amount of time and results of the evaluation may vary between analyzers. Artificial intelligence (AI) is widely used for image segmentation, but was not applied for the analysis of such images before.
Aims: We aimed to develop an AI-based, automatic IS analysis software to reduce image analysis time and increase analysis reproducibility.
Methods: Images were acquired from experiments on male Wistar rats with 30 minutes occlusion and 120 minutes reperfusion of the left anterior descending coronary artery. Area-at-risk (AAR) was negatively stained with Evans Blue, and IS was negatively stained with TTC. A U-NET architecture-inspired neural network with five downscaling and five upscaling layers was implemented and trained using cross-entropy loss and Adam optimizer. AI was trained on 1200 randomly selected slices and then results were compared with human expert analyzer.
Results: Trained AI labelled the myocardium optimally and the analysis error on IS/AAR was comparable to the variance of the manual analysis. When slices were pre-annotated with the AI and then a human analyzer applied adjustments, changes were moderate in IS/AAR values of the animals (9.41±8.27 %; mean±SD), however, the mean and SEM of the individual experimental groups did not significantly differ.
Conclusion: Since the error of our AI software was comparable to the variance between human analyzers, and the modifications deemed necessary by human analyzer were moderate after pre-annotation of images, our AI software can greatly reduce measurement time and may increase data reproducibility. However, it may not be applicable as an automated, human-free analysis software in its current version.
Funding: Project was supported by Thematic Excellence Programme (2020-4.1.1.-TKP2020) of the Ministry for Innovation and Technology in Hungary, within the framework of the Therapeutic Development and Bioimaging thematic programmes of the Semmelweis University, NKFIH of Hungary K139237 to AG. C.K. was supported by NTP-NFTÖ-21-B-0300 and SE250+ Excellence Scholarship (EFOP-3.6.3-VEKOP-16-2017-00009), NRDI Fund (2019-1.1.1-PIACI-KFI-2019-00367)