ResNet to the ResCue: An automated approach for the detection of measurable residual disease in patients with acute myeloid leukemia.
Type of the data | Collection | |
Total size of the dataset | 1070315463 | |
Author | Thielecke, Lars | |
Author | Roehnert, Maximilian-Alexander | |
Upload date | 2026-03-18T12:28:04Z | |
Publication date | 2026-03-18T12:28:04Z | |
Data of data creation | 2025 | |
Publication date | 2026-03-18 | |
Abstract of the dataset | This repository provides a minimal, end‑to‑end example pipeline demonstrating how a trained ResNet‑34 model can recognize MRD‑associated patterns in UMAP embeddings derived from flow‑cytometry data of AML patients. ResNet‑based classifiers are widely used for image‑recognition tasks , making them well‑suited for distinguishing subtle MRD‑related patterns in UMAP‑transformed cytometry data. The pipeline consists of a small collection of Python and R scripts organized as a lightweight workflow, following the standard idea of pipelines as sequences of data‑processing and prediction steps . Included in the repository are: - a single example patient dataset (raw data, pre‑processed data, and generated image) - an R script for preprocessing the raw flow‑cytometry data - a Python script that converts the tabular preprocessed data into 2D images - the representative UMAP embedding needed for generating standardized 2D representations of patient-specific data - a Python script that loads the ResNet‑34 architecture (including the custom classifier head), initializes the trained weights , and runs the prediction procedure to distinguish MRD‑positive from MRD‑negative image patterns This example is intentionally minimal: it is not a production‑ready pipeline but an educational demonstration of how the core steps—data preparation, image generation, and model prediction—link together in a transparent, reproducible workflow. | |
Public reference to this page | https://opara.zih.tu-dresden.de/handle/123456789/2141 | |
Public reference to this page | https://doi.org/10.25532/OPARA-1134 | |
Publisher | Technische Universität Dresden | |
Licence | Attribution-NonCommercial-NoDerivatives 4.0 International | en |
URI of the licence text | http://creativecommons.org/licenses/by-nc-nd/4.0/ | |
Specification of the discipline(s) | 2::22::205::205-14 | |
Specification of the discipline(s) | 2::22::205::205-07 | |
Title of the dataset | ResNet to the ResCue: An automated approach for the detection of measurable residual disease in patients with acute myeloid leukemia. | |
Project abstract | The detection of Measurable Residual Disease (MRD) in patients with Acute Myeloid Leukemia (AML) during and after treatment is an important diagnostic procedure to prospectively identify leukemia and adjust patient therapy. Besides a range of molecular techniques, multiparametric flow cytometry (MFC) is a pivotal method for MRD detection, but the manual evaluation of MFC-MRD data is time-intensive, subjective and difficult to standardize across multiple centers. To facilitate broader accessibility and provide objective diagnostic support, we propose a machine learning–based approach for AML-MRD detection based on raw MFC measurements. Unlike classical methods, which rely on two-dimensional manual gating of selected antigen combinations, our approach uses the full multi-dimensional structure of MFC data to identify features of aberrant cell populations. The data library comprised 1,361 samples: 1,271 from AML patients (765 at diagnoses, 336 MRD-positive follow up and 170 MRD-negative follow up samples) and 90 control measurements from non-AML individuals, with single-center expert-derived MFC diagnoses serving as ground truth. Utilizing the UMAP-algorithm, we translated raw MFC data into 2D images which served as training data for a ResNet34 model. After training, we achieved a mean sensitivity of 0.91, a specificity of 0.90, and an F1-score of 0.94. We validated our methodology by comparing its performance to expert-derived manual diagnostics and assessed its potential as a standalone decision-support tool for MRD detection in AML patients. We demonstrate that a machine learning-based approach is feasible even for MRD detection in AML patients. Our method is able to determine MRD status with a level of accuracy comparable to the consensus among human experts analyzing the same patient samples. | |
Funding Acknowledgement | This work was supported by Wilhelm Sander Stiftung (MinimaL, #2021.035.1), Deutsche Forschungsgesellschaft (PERDAM, #318488004), Technische Universität Dresden (MeDDrive, intramural funding, #60466 and #60534), BMFTR (TRACK, #01KD25030), CAMINO, the Advanced Clinician Scientist Program of the TU Dresden, funded by BMFTR (01EO2101) and the Else Kröner Fresenius Center for Digital Health. | |
Project title | ResNet to the ResCue: An automated approach for the detection of measurable residual disease in patients with acute myeloid leukemia. |
Collections
