AI Automation of Flow Cytometry to Identify Myeloma Restricted; Files Only

Abstract

Deep learning to automate flow cytometry interpretations for disease diagnosis is an underexplored area, but it is an important direction for the medical field to offer pathologists a second opinion, catch myeloma early for better patient outcomes, and reduce the workload of all medical professionals involved in the data processing and interpretation of flow cytometry results. Convolutional neural networks (CNNs) and attention-based multiple instance learning (ABMIL) models have show promising results for flow cytometry interpretation. However, models continue to struggle with the subtlety and nuance of malignant plasma cell populations. We present a novel approach based on an ABMIL pipeline called Branched Attention for Monoclonal Sensitivity (BAMS) to navigate the large volume of flow cytometry data through several attention branches that reduce dimensionality by identifying special cell populations before the final classification. We hypothesize that reducing the context from a large population of events to a smaller population of interest through sequential branching ”sub-problems” will allow the model to be more sensitive to the abnormal plasma cells. Our model is evaluated on our own new dataset, MM25, which contains data from patients at Emory University Hospital. Our approach outperforms all other models tested, and it makes fewer false negative predictions compared to all other models tested. To the best of our knowledge, we are the first work to take a deep learning multi-branch attention approach to flow cytometry interpretation and the first work to study the application of deep learning models to myeloma flow cytometry.

Table of Contents

Contents

1 Introduction 1

1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1.1 Myeloma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1.2 Flow Cytometry . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.4 Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.5 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Background 6

2.1 Previous Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.1.1 Classical Machine Learning . . . . . . . . . . . . . . . . . . . 6

2.1.2 Convolutional Neural Networks . . . . . . . . . . . . . . . . . 7

2.1.3 Attention-based Multiple Instance Learning . . . . . . . . . . 7

2.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

3 Approach 11

3.1 Preprocessing and Models . . . . . . . . . . . . . . . . . . . . . . . . 11

3.1.1 Classical Machine Learning . . . . . . . . . . . . . . . . . . . 11

3.1.2 Convolutional Neural Networks . . . . . . . . . . . . . . . . . 13

3.1.3 Attention-based Multiple Instance Learning . . . . . . . . . . 13

3.1.4 Branched Attention for Monoclonal Sensitivity (BAMS) . . . . 13

3.2 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.2.1 Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3.2.2 Sourcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4 Experiments 20 4.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.2 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.3.1 Ablation Studies . . . . . . . . . . . . . . . . . . . . . . . . . 22

5 Analysis 24

5.1 Comparison with ABMIL . . . . . . . . . . . . . . . . . . . . . . . . 24

5.2 Error Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

5.3 Fine-tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

5.4 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

6 Conclusion 29

Bibliography 30

About this Honors Thesis

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Emory College
Department	Computer Science
Degree	B.S.
Submission	Honors Thesis
Language	English
Research Field	Computer Science Biology, Cell Biology, Microbiology
Keyword	Flow Cytometry Diagnosis Cancer Automation Myeloma AI
Committee Chair / Thesis Advisor	Kai Shu, Emory University
Committee Members	Carl Yang, Emory University Youngjoong Kwon, Emory University

Last modified

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	File download under embargo until 28 November 2026	2026-04-12 06:37:45 -0400	File download under embargo until 28 November 2026

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions