Incorporating Contextual Information in White Blood Cell Identification

Xubo B. Song, Yaser S. Abu-Mostafa, Joseph Sill, Harvey KasdanNeurIPS 1997

TL;DR

This paper challenges the standard machine learning assumption that data points are independent and identically distributed (i.i.d.). By incorporating "context"—specifically, the identities of surrounding cells in a blood sample—the authors significantly reduce false alarm rates in white blood cell classification. Their approach improves the false positive rate for specimen normality from roughly 50% down to just 10%.

What Problem Does This Paper Address?

In traditional pattern recognition, we usually treat each object as an isolated island. If you are classifying images of white blood cells (WBCs), the standard approach is to look at one cell, decide what it is, and move to the next.

Examples of various white blood cell types, including lymphocytes and blasts.

However, the authors point out that in hematology, abnormal cells are much more likely to appear in bunches than in isolation. If a single cell looks slightly like a "blast" (a clinically significant immature cell) but is surrounded by 600 perfectly healthy cells, it is likely just a noisy measurement or an unusually shaped healthy lymphocyte. Conversely, if that same cell is surrounded by other borderline-abnormal cells, the probability that it is genuinely a blast skyrockets.

The gap the authors tackle is how to mathematically formalize this "intuition of the company it keeps" to improve classification accuracy and, more importantly, reduce the high cost of false alarms in clinical settings.

Methods

The authors propose a two-stage framework to integrate this contextual information:

  1. Cell-by-Cell Classification: A standard feed-forward neural network (20 inputs, 15 hidden units, 14 outputs) processes 20 extracted features—such as cell area, edge pixels, and color distribution metrics—to produce a "no-context" probability for each cell.
  2. Contextual Adjustment: They introduce a Context Ratio based on Bayes' rule. This ratio captures the dependence between objects by looking at the joint distribution of all classes in a specimen.

The Computational Challenge: Classifying $N$ objects into $D$ classes normally requires evaluating $D^N$ combinations. With 14 cell types and 600 cells per specimen, $14^{600}$ is computationally impossible to calculate directly.

The Innovation: The authors simplify the problem using domain-specific observations. They note that clinical concern primarily focuses on "blasts," and blasts are almost exclusively confused with "lymphocytes." By assuming all other cell types are correctly identified by the initial neural network, they reduce the search space to a recursive calculation that only considers the $K$ most likely blast candidates. This turns an exponential problem into a manageable linear one.

Datasets

The study utilized a dataset provided by IRIS, Inc., consisting of:

  • Specimens: 220 blood samples collected at Harbor UCLA Medical Center.
  • Total Cells: 13,200 individual cell images (roughly 600 images per specimen).
  • Processing: Cells were dyed with Basic Orange 21 and captured via flow microscopic imaging. Each image is roughly 25x25 pixels.

Key Insights and Ideas

  • Breaking i.i.d.: The core insight is that the identity of one cell provides a prior for the identity of the next. In biological systems, pathology is rarely a localized, single-cell event; it is a systemic shift in distribution.
  • The "Context Ratio": The authors define context as the ratio of the joint probability of all labels to the product of their individual probabilities. If cells are independent, this ratio is 1. If they are dependent, this ratio acts as a "correction factor" for the initial classification.
  • Clinical Utility vs. Raw Accuracy: Interestingly, the overall cell-level classification accuracy only improved by 1% (from 88% to 89%). However, the specimen-level diagnosis improved drastically. This highlights that in medical AI, the goal isn't just "more correct pixels," but more reliable clinical decisions.

Key Results

The impact of adding context was most visible in the reduction of "false alarms" (healthy samples incorrectly flagged as abnormal):

Performance comparison showing the dramatic reduction in false positives when contextual information is used.

By using context, the system eliminated the vast majority of cases where a few noisy healthy cells would have triggered an expensive and stressful manual review by a pathologist. Crucially, this massive reduction in false positives (~50% down to ~10%) was achieved without missing a single true positive (maintaining a 0% false negative rate).

Limitations and Future Work

The primary limitation of this approach is its dependency on the simplification heuristic. The current algorithm assumes that the only significant ambiguity is between blasts and lymphocytes. If a new type of pathology appeared where three or four different cell types were frequently confused, the current recursive simplification might not hold, and the computational complexity would creep back up.

The authors suggest that this framework could be generalized to other fields where context is king, such as speech recognition (where the meaning of a phoneme depends on the words around it) and character recognition (where a blurry letter is identified by the word it inhabits).

Summary updated

BibTeX
@inproceedings{song1997incorporating,
  author = {Xubo B. Song and Yaser S. Abu-Mostafa and Joseph Sill and Harvey Kasdan},
  title = {{Incorporating Contextual Information in White Blood Cell Identification}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1997},
}
Figure