AutoDBC : Scalable System for Classification of White Blood Cells from Leishman Stained Blood Stain Images

Published: 2012-11-17 17:06:03

Categories: Robotics & Perception Systems

<- Back to Blog Home

This page originally pointed to the old project post on the AutoDBC WordPress site.
To keep the content local, here is a short summary of the work.

Overview

AutoDBC is a pipeline for automated differential white blood cell (WBC) counting from Leishman-stained peripheral blood smear images. The core motivation was to reduce manual annotation time and improve consistency in routine hematology workflows.

Method Summary

The system follows four stages:

  1. Stain normalization to improve consistency across slide images.
  2. Segmentation of nucleus and cytoplasm (including active contour based refinement).
  3. Feature extraction, including a custom "number of lobes" feature for subtype discrimination.
  4. Multiclass classification using Naive Bayes with Laplacian correction and incremental-learning-oriented workflow.

Figures from Original Post

Segmentation Scheme Segmentation Scheme (from original AutoDBC post, rehosted image URL captured in the WordPress export snapshot).

Feature Extraction and Classification Feature Extraction and Classification (from original AutoDBC post, rehosted image URL captured in the WordPress export snapshot).

Reported Results

The published article reports overall accuracy around 92% on both training and testing splits.

Credits and Sources