Deep learning improves the identification of neutrophil abnormalities in immune and inflammatory conditions

Introduction

The identification of inflammatory and immunological diseases, which affect millions of people, requires accurate and early diagnosis to optimize treatment. However, these diagnoses often rely heavily on visual analysis by expert clinical pathologists, delaying timely intervention.

Neutrophils, the most important immune cells, are critical in defending against infections and regulating the inflammatory response. While conventional morphological analysis systems can identify normal neutrophils, they have difficulty detecting specific alterations, such as those seen in bacterial infections, severe inflammation, and autoimmune disorders. This limitation poses a significant challenge to timely and accurate diagnosis.

Objective

This work aims to develop an automated deep-learning-based system to differentiate normal neutrophils from those with abnormalities characteristic of various pathologies, including bacterial infections, severe inflammation, and autoimmune disorders.

Methodology

The images were obtained at the Core Laboratory of the Hospital Clínic de Barcelona using the Cellavision DM96 morphological analysis system. Pathologists validated 5,492 images: normal neutrophils (4,595), hypogranulated neutrophils (494), and neutrophils with inclusions (Döhle bodies: 139, cryoglobulins: 191, bacteria: 73).

To address imbalance, the Pareto rule was applied, starting with the smallest group (bacteria), generating 138 training images and oversampling to 276. This value balanced each neutrophil class in training and proportionally divided the dataset into training (828), validation (216), and test (4,622) sets. Data Augmentations (rotation, zoom, mirroring) were applied. Two ResNet152-based models A and B classify general categories and inclusions.

Results

The deep learning system showed high accuracy: 99% in Model A and 85% in Model B for classifying normal neutrophils and those with inclusions. It effectively identified normal, hypogranulated neutrophils, and those containing bacteria, cryoglobulins, and Döhle bodies, demonstrating its clinical value.

Conclusion

The proposed system effectively identifies normal neutrophils and those linked to bacterial infections, severe inflammation, and autoimmune disorders, showing potential for enhancing hematological disease diagnosis.