Identifying image features that associate strongly with diagnostic or prognostic classes in large-scale, multi-channel spatial imaging is challenging without pixel-level annotations. We present DREAMER-S, an attention-based multiple-instance learning (MIL) framework that, using only image- or slide-level labels, learns spatial features within 3D imaging hypercubes that are most informative for downstream classification. We demonstrate DREAMER-S on Quantum Cascade Laser infrared (QCL-IR) tissue imaging, where attention weights are rendered spatially to highlight class-relevant spectral instances without manual annotation. Because the MIL attention layer assigns interpretable importances to spatial instances, the method is broadly transferable to spatial-biology applications that require instance-level filtering to focus towards salient regions of interest in high-content datasets. We further evaluate DREAMER-S on a chemotherapy-response task in a colorectal cancer patient-derived xenograft (PDX) model. After tuning, DREAMER-S separated spectral instances from a chemo-sensitive PDX (CRC0344) and a less responsive PDX (CRC0076) with an F1 score of ~0.95. To validate explainability, we linked model saliency to cellular physiology, observing that, (i) unsupervised UMAP embeddings of high-attention spectra stratified samples by treatment (chemotherapy, apoptosis sensitizer, combination, vehicle), and (ii) selected spectral markers correlated with pro-apoptotic proteins measured independently in the same PDX system. Together, these results support a mechanistic link between spectral signals and apoptosis pathways and position DREAMER-S as an efficient, interpretable approach for analysing high-content spatial-biology imaging datasets.
Copyright: © 2026 Rafsanjani et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.