Dual-region deep learning for odor discrimination from local field potentials

Decoding odor information in the mammalian brain is a dynamic process involving interconnected brain regions, such as the Olfactory Bulb (OB) and Piriform Cortex (PCx). Isolating the functional contribution of each region to odor discrimination is challenging due to the complexity of simultaneous multi-region recordings and recurrent feedforward/feedback pathways. To address this, we introduce the Dual-Region Network (DRN), a novel deep learning architecture designed to fuse Welch-derived power spectra (1–600 Hertz) from local field potentials (LFP) recordings in the OB and PCx. The DRN combines residual layers with an attention-based fusion mechanism, providing a computationally lightweight, end-to-end decoding framework. By evaluating the DRN on a public dataset, we achieved 90% accuracy for four odors and 73% for seven odors, significantly outperforming re-implemented deep learning baselines (Electroencephalogram Network (EEGNet): 62.50%, Shallow Convolutional Network (Shallow ConvNet): 62.40%) and approaching the performance of invasive single-unit decoders—without requiring spike sorting. Crucially, LFP-based decoding is advantageous in spatially segregated regions: unlike single-unit responses, which vary with electrode placement relative to active glomeruli, LFPs capture net population dynamics. This yields a robust encoding of odor identity, largely independent of the precise recording location or the quality of unit isolation. Our model performs unsupervised selection of stimulus-selective frequency components, leading to high odor-discrimination power. The ablation simulation confirms that odor molecule encoding is strongly influenced by odor delivery time and region-specific frequency components. The DRN offers a reproducible approach to dual-region neural decoding, providing a transferable framework for dissecting distributed computations within brain circuits. • A novel dual-region network decodes odors from OB-PCx field potentials. • Achieves 90% accuracy, rivaling invasive single-unit decoders. • Odor identity is encoded within one second of inhalation onset. • OB high-gamma and PCx low-frequency bands show complementary coding. • Dual-region fusion outperforms single-region baselines.