Learning Channel Importance for High Content Imaging with Interpretable Deep Input Channel Mixing

TL;DR

The paper tackles the challenge of interpretable channel importance in multi-channel high-content imaging by introducing DCMIX, a lightweight, end-to-end trainable image blending layer that condenses multiple spectral channels into a single informative representation. By learning nonnegative channel weights, DCMIX provides phenotype-focused interpretations while preserving competitive classification performance on MNIST and RXRX1, and does so with markedly lower model complexity and faster runtimes than attention-based baselines. The authors demonstrate strong alignment with ground-truth channel importance (Spearman ρ ≈ 0.89) and show practical advantages in real-world datasets, highlighting scalability to many channels and applicability beyond biomedicine. Overall, DCMIX offers a scalable, interpretable fusion mechanism for multi-spectral imaging that preserves performance and enables rapid, channel-specific biological insights.

Abstract

Uncovering novel drug candidates for treating complex diseases remain one of the most challenging tasks in early discovery research. To tackle this challenge, biopharma research established a standardized high content imaging protocol that tags different cellular compartments per image channel. In order to judge the experimental outcome, the scientist requires knowledge about the channel importance with respect to a certain phenotype for decoding the underlying biology. In contrast to traditional image analysis approaches, such experiments are nowadays preferably analyzed by deep learning based approaches which, however, lack crucial information about the channel importance. To overcome this limitation, we present a novel approach which utilizes multi-spectral information of high content images to interpret a certain aspect of cellular biology. To this end, we base our method on image blending concepts with alpha compositing for an arbitrary number of channels. More specifically, we introduce DCMIX, a lightweight, scaleable and end-to-end trainable mixing layer which enables interpretable predictions in high content imaging while retaining the benefits of deep learning based methods. We employ an extensive set of experiments on both MNIST and RXRX1 datasets, demonstrating that DCMIX learns the biologically relevant channel importance without scarifying prediction performance.

Figures (2)

• Figure 1: Blue arrows denote steps and gray boxes actions in our workflow, respectively. In the first step (1.), we take a multi-channel cellular image and split it into single channels. Subsequently, we mix the channel within our DCMIX layer to obtain the most important part of each channel. In the second step (2.), we take the blended image into our classification network.
• Figure 2: Visualization of Spearman's rank correlation coefficient of the channel importance estimates for all different methods from Table \ref{['tab1']}. A value of -1 indicates maximal ranking difference between the channel importance estimates, 1 indicates no difference. Matrix has been sorted using average linkage hierarchical clustering with euclidean distance.