Automated analysis of flow cytometry using deep learning for the detection of B-cell neoplasms

Abstract

B-cell neoplasms are the most prevalent type of non-Hodgkin lymphoma, including a diverse and heterogenous group of entities. Immunophenotyping with a high-throughput technology like flow cytometry is a standard diagnostic procedure in evaluating B-cell neoplasms. While multi-parameter flow cytometry (FCS) has become a cornerstone in clinical decision-making for leukemia and lymphoma, the data analysis requires manual gating of cell populations, which is time-consuming, subjective, and often limited to a two-dimensional space. In recent years, machine learning has become a popular approach for automating manual gating. Many automated gating algorithms require expert supervision and cannot classify the data into diagnosis labels. Furthermore, these algorithms still limit the analysis to a two-dimensional space, leading to the loss of information in the high-dimensional FCS data.

We hypothesize that the wealth of information captured in “n”-dimensional FCS data can be analyzed by current computer vision methods when represented as image data. We, therefore, transformed FCS raw data into a multicolor low-resolution image using self-organizing maps. These images are then analyzed and classified using a convolutional neural network. By this means, we built an artificial intelligence (AI) that not only can distinguish diseased from healthy samples but also differentiate seven subtypes of mature B-cell neoplasm. We trained our model with 18,274 cases, including chronic lymphocytic leukemia and its precursor monoclonal B-cell lymphocytosis, marginal zone lymphoma, mantle cell lymphoma, prolymphocytic leukemia, follicular lymphoma, hairy cell leukemia, lymphoplasmacytic lymphoma and achieved a weighted F1 score of 0.94 on a separate test set of 2,348 cases.

Next, we extend our AI model to multiple datasets and FCS panels using transfer learning in conjunction with FCS data merging. We demonstrate how transfer learning can be applied to boost the performance of models with much smaller datasets acquired with different FCS panels. We trained a new AI for four additional datasets by transferring the features learned from our base model. Our workflow increased the model's overall performance and, more prominently, improved the learning rate for small training sizes.