AI-assisted diagnosis of anemia through peripheral smear image analysis: A cross-validation study

TL;DR

This paper presents an AI model that accurately detects and classifies anemia from blood smear images, showing strong agreement with expert diagnoses.

Contribution

A deep semi-supervised learning model for anemia diagnosis using only 25% annotated data, achieving high accuracy and F1-scores.

Findings

The model achieved 93.4% classification accuracy and F1-scores above 90% for key anemia subtypes.

It showed strong agreement with expert diagnoses (κ = 0.89) and performed well in detecting microcytic and sickle cell anemia.

The model significantly reduced diagnostic time and is suitable for resource-limited settings.

Abstract

A deep semi-supervised learning model for automating anemia detection and classification from peripheral blood smear images is of
interest. A convolutional neural network was trained on 3,200 images, with only 25% annotated by expert hematologists. The model
achieved a classification accuracy of 93.4% and F1-scores above 90% for key anemia subtypes, demonstrating strong agreement with expert
diagnoses (κ = 0.89). It significantly reduced diagnostic time and performed well in detecting microcytic and sickle cell anemia.
This AI-based framework shows great potential for accurate anemia diagnosis, especially in resource-limited settings.