One-Blood

Introduction

In the field of Blood-based disease diagnosis, identifying blood cells subtypes of patients is significant. Human have been doing the job of identifying blood cells subtypes. One Blood aims to devise to automate this job by introducing a machine learning model that does the classification job.

Goal

We aim to develop a machine learning model that can predict the blood cell subtype of a blood cell image.

Architecture

Input

A blood cell iamge

Convolutional Neural Network

• Pretrained CNN (Inception v3 or VGG 19)
• Custom CNN

Dense Network

Number of hidden layers

Work in progress

Number of hidden units per layer

Work in progress

Dropout value

Work in progress

Final Layer

• Loss function: Categorical Cross-Entropy
• Number of hidden units: 4

Output Labels

• Eosinophil
• Lymphocyte
• Monocyte
• Neutrophil

Model Optimizer

Adam or RMSprop

Expected Result

The model will be used to aid in prediction of the subtype of blood cells as a supplement to or in place of medical experts. We expect the model to predict blood cell subtypes at an accuracy higher than 80%.

Expected Challenges

We expect that naturally there is human error within the data already because the images of the blood cells are originally labeled by human doctors, but seeing as how they are the best-versed in the topic, we are willing to accept this level of accuracy.

References

[1] https://www.kaggle.com/paultimothymooney/blood-cells