Interpretable Embeddings for Segmentation-Free Single-Cell Analysis in Multiplex Imaging

TL;DR

This work proposes a segmentation-free deep learning approach that leverages grouped convolutions to learn interpretable embedded features from each imaging channel, enabling robust cell-type identification without manual feature selection.

Abstract

Multiplex Imaging (MI) enables the simultaneous visualization of multiple biological markers in separate imaging channels at subcellular resolution, providing valuable insights into cell-type heterogeneity and spatial organization. However, current computational pipelines rely on cell segmentation algorithms, which require laborious fine-tuning and can introduce downstream errors due to inaccurate single-cell representations. We propose a segmentation-free deep learning approach that leverages grouped convolutions to learn interpretable embedded features from each imaging channel, enabling robust cell-type identification without manual feature selection. Validated on an Imaging Mass Cytometry dataset of 1.8 million cells from neuroblastoma patients, our method enables the accurate identification of known cell types, showcasing its scalability and suitability for high-dimensional MI data.

Figures (5)

• Figure 1: A: Core unit of the proposed architecture. White blocks represent convolutional operations, formatted as: input features, $\text{kernel size} \times \text{kernel size}$, and output features. Here, $D$ is the number of input features, $G$ denotes the number of groups, and $E$ indicates the expansion factor of the block. B: Example data flow through the model, starting from a $32 \times 32 \times C$ pixel input patch, where $C$ is the number of image channels. White blocks represent convolutional layers with a $3 \times 3$ kernel size and number of groups ($G$). The interpretability stage (green) is used to obtain channel-specific activation, enabling expert interpretation. The final representation is generated by processing the previously disentangled channels together, allowing for crosstalk between them. C: Markers used in the IMC panel.
• Figure 2: UMAP plots showing the contribution of specific markers to the representation space for selected channels: CD20 (expressed on B cells), GD2 (expressed on neuroblastoma cells), CD3 (expressed on T cells), CD8 (expressed on cytotoxic T cells), CD4 (expressed on helper T cells) and GZMB (indicative of cytotoxic T cells and natural killer cells).
• Figure 3: A: Heatmap of average activations from the interpretability stage of our proposed method for each Phenograph cluster across all 34 markers in the panel. B: UMAP embedding ofthe representation stage of our method, colored by cell type based on benchmark phenotyping using nucleus segmentation. C: Same UMAP as in (B), but with Phenograph clusters and their cell type assignments as identified by domain experts using our approach.
• Figure 4: Rediscovery rate of each cell type cluster identified in the benchmark analysis using segmentation, compared to the cell type clusters assigned by domain experts using our proposed method. Normalized by row.
• Figure 5: A: Detailed UMAP visualization highlighting distinct T cell subclusters identified by Phenograph, showing the activation obtained from the interpretability stage of our proposed approach for key T cell markers: CD3, GZMB, CD8, and CD4. B: Average activation per subcluster for all 34 markers in the panel. C: Gallery of representative cells from each of the 7 subclusters, with markers displayed in distinct colors.