Introduction and objectives:

Myelodysplastic syndrome (MDS) is a blood disorder marked by abnormal blood cell production and a high risk of leukemia. Neutrophils, a type of leukocyte, can exhibit hypogranulation, an early indicator of MDS. While automated analyzers can identify leukocyte types, they cannot distinguish between normal and hypogranulated neutrophils, a challenge that deep learning techniques can address.

The objective of this work is to develop a two-stage identification system using convolutional neural networks (CNNs). The first stage classifies five types of leukocytes, including normal and hypogranulated neutrophils, while the second stage distinguishes between normal and hypogranulated neutrophils.

Methods

Images were collected at the Hospital Clínic de Barcelona using the CellaVision DM96 analyzer, resulting in a dataset of 500 images of basophils, lymphocytes, eosinophils and monocytes and 5089 images of neutrophils (4595 normal and 494 hypogranulated). The dataset was split into training sets (320 images per leukocyte type, with 160 normal and 160 hypogranulated neutrophils), validation and test sets, and the rest were used for the test set. Two CNN-based models were developed: the first one with VGG19 architecture to differentiate between leukocyte types and the second one with ConvNeXt architecture to distinguish normal from hypogranulated neutrophils. A final proof of concept was performed with 1000 images of neutrophils from MDS patients to simulate real clinical conditions.

Results

The test set results showed 99.57% accuracy for the first model and 99.32% for the second, identifying 4415 normal and 282 hypogranulated neutrophils. In the proof of concept, the accuracies were 96.7% for the first model and 80% for the second, with sensitivities of 87.9% for hypogranulated neutrophils.

Conclusion

The proposed system showed high accuracy in testing and acceptable performance in the proof of concept for detecting neutrophil hypogranulation in MDS patients. However, its clinical performance was lower, indicating a need for more diverse data and refinement for improved accuracy in real-world applications.