Bright 4B: Scaling Hyperspherical Learning for Segmentation in 3D Brightfield Microscopy

TL;DR

Bright-4B addresses the challenge of 3D segmentation from label-free brightfield data by learning on the unit hypersphere with a 4B-parameter transformer backbone. It introduces a geometry-faithful anisotropic patch embedding, Native Sparse Attention that fuses local, coarse, and selective global context, and a slot-based Soft MoE with Dynamic HyperConnections to stabilize training and adapt capacity. The model demonstrates morphology-accurate segmentation of nuclei, mitochondria, and other organelles directly from brightfield stacks, outperforming CNN and Transformer baselines while maintaining a feasible 26 GB inference footprint on a single GPU. This work sets a foundation for scalable, label-free 3D cell mapping and provides code, pretrained weights, and models to advance microscopy AI.

Abstract

Label-free 3D brightfield microscopy offers a fast and noninvasive way to visualize cellular morphology, yet robust volumetric segmentation still typically depends on fluorescence or heavy post-processing. We address this gap by introducing Bright-4B, a 4 billion parameter foundation model that learns on the unit hypersphere to segment subcellular structures directly from 3D brightfield volumes. Bright-4B combines a hardware-aligned Native Sparse Attention mechanism (capturing local, coarse, and selected global context), depth-width residual HyperConnections that stabilize representation flow, and a soft Mixture-of-Experts for adaptive capacity. A plug-and-play anisotropic patch embed further respects confocal point-spread and axial thinning, enabling geometry-faithful 3D tokenization. The resulting model produces morphology-accurate segmentations of nuclei, mitochondria, and other organelles from brightfield stacks alone--without fluorescence, auxiliary channels, or handcrafted post-processing. Across multiple confocal datasets, Bright-4B preserves fine structural detail across depth and cell types, outperforming contemporary CNN and Transformer baselines. All code, pretrained weights, and models for downstream finetuning will be released to advance large-scale, label-free 3D cell mapping.

Figures (15)

• Figure 1: Bright 4B is a 4B-parameter foundation model for segmenting subcellular structures in 3D brightfield microscopy images that learns on the unit hypersphere. It accepts label-free 3D brightfield microscopy images as input and predicts subcellular structures in 3D, preserving fine structural detail across depth and cell types, without requiring explicit fluorescence images.
• Figure 2: Anisotropic 3D patch embedding. A PSF-aware axial low-pass with stride $p_z$ prevents aliasing, followed by tiling in the $xy$ plane ($16\times16$). A bias-free projection maps each tile to $h$ dimensions, producing a geometry-faithful token lattice $(B,D',H',W',h)$ with $D'=D/2$. This $(2\times16\times16)$ design respects confocal anisotropy and supports accurate 3D segmentation.
• Figure 3: Hypersphere Setup.[A] The pre-update $x_0$ and proposal $\hat{u}$ live on the unit sphere. We use LERP + $L_2$ normalization (chord + radial projection), approximating SLERP for small angles. [B] Sublayer outputs from MoE or Attention take small, feature-wise steps that remain stable on the hypersphere. [C] A two-input mixer with gate $\beta$ blends in $\mathbb{R}^d$; the depth gate arbitrates new evidence versus prior memory per feature. [D]$E$ streams are processed in parallel, then reduced and normalized. Intuitively, width = multiple latent opinions, merge = consensus on the hypersphere.
• Figure 4: Bright 4B Architecture. Each block follows PreNorm-Attention-Residual and PreNorm-FFN-Residual but with two additions: a cross-stream mixer that blends $E$ residual streams, and a spherical residual gate that applies unit-norm LERP updates. We replace MHA with Native Sparse Attention (sliding window, compressed memory, global selection) and the MLP with SoftMoE (slot-based soft routing + load balancing). Shapes are preserved $(B,N,d)$ throughout.
• Figure 5: For each model, we visualize the voxel-wise difference between the pseudolabel and the predicted mask.