Immune response is a critical indicator to assess fish health. The profile of white blood cells (WBCs) is a valuable metric for evaluating immune status, as variations in WBC abundance can reflect underlying health issues. In this project, we employed machine learning (ML) to automate image recognition of WBCs in blood smears from juvenile largemouth bass. The goal is to replace manual WBC counting, which is time-consuming, requires technical training, and is potentially inconsistent among readers.
Blood smear slides were created from 38 fish. Three independent readers manually counted and classified 100 WBCs per slide. These slides were scanned at 83x, digitized using an Aperio ScanScope CS, and uploaded for labeling using SageMaker Ground Truth software. “Regions of interest” (ROIs) were selected and divided into 48 “tiles”. WBCs were labeled to create a labeled image dataset to train the model (Fig. 1). Labeled cells included 1242 lymphocytes, 493 monocytes and 297 granulocytes; 90% of these were used to train the model and 10% were used to test the model.
To validate the ML model, we compared manual cell counts among the 3 human readers, manual cell counts to automated cell counts by the model, and automated cell counts of novel tiles taken from the training slides. Manual counts of lymphocytes were most similar among the three readers, with each reader being within 95-106% of the average count for all three. Manual counts of granulocytes and monocytes were more variable, with each reader within 73-149% of the average. These manual cell counts will be compared to automated counts to evaluate effectiveness of the model. At present, the model is approximately 80% accurate in its ability to correctly recognize all three cell types in blood smears from largemouth bass. Additional training tiles can be labeled to increase model accuracy. The potential benefits of this automated tool include increased accuracy and efficiency in WBC analysis, less reliance on specialized training, and enhanced accessibility.