| | --- |
| | {} |
| | --- |
| | |
| | # ATLAS-1: 3D Drug-Treated Cancer Cells |
| |
|
| | **ATLAS-1** is a research dataset of single-cell **3D morphologies** captured under different drug treatments. |
| | It contains **1,500** WM266-4 human melanoma cells embedded in **collagen** and imaged using **oblique-plane light-sheet microscopy**. Each cell is released as a **1024-point point cloud** sampled from a reconstructed **surface mesh**. |
| |
|
| | - **Total cells:** 1,500 |
| | - **Classes / conditions:** |
| | - `nocodazole` (500) — microtubule inhibitor; rounded morphology |
| | - `blebbistatin` (500) — myosin II inhibitor; elongated/spindle morphology |
| | - `control` (500) — untreated baseline variability |
| | - **Representations:** point clouds (1024 pts/cell) and watertight meshes |
| | - **Intended use:** benchmarking point-cloud models on real biological shapes; studying drug-induced 3D morphology |
| |
|
| | > This dataset accompanies the **PointMIL** project (inherently interpretable point-cloud classification via MIL). |
| |
|
| | --- |
| |
|
| | ## Dataset Summary |
| |
|
| | Cells were segmented, meshed with **marching cubes** (+ Laplacian smoothing), then **uniformly sampled** to 1024 points per cell from the surface mesh. Files are organized by condition with a metadata CSV that holds labels and basic experiment info. |
| |
|
| | **Why ATLAS-1?** |
| | - Real, noisy, heterogeneous 3D cell shapes (not synthetic CAD) |
| | - Clear treatment effects (rounded vs elongated) for interpretable benchmarks |
| | - Comes with both **mesh** and **point-cloud** views for flexible pipelines |
| |
|
| | --- |
| |
|
| | ## Repository Layout |
| |
|
| | 🚀 PointMIL is an ICCV 2025 Highlight — powering interpretable 3D cell shape analysis at Sentinal4D |
| |
|
| |
|
| |
|
| | At Sentinal4D, we’re pioneering the analysis of 3D cell shapes, and we believe models should be trustworthy by design. |
| |
|
| |
|
| |
|
| | PointMIL is our unified Multiple Instance Learning (MIL) framework that delivers state-of-the-art accuracy across 3D biomedical datasets and inherent interpretability. |
| |
|
| |
|
| |
|
| | Why this matters for biology |
| |
|
| | - Mapping drug effects in 3D: See which cell regions the drug affects, e.g., rounding vs. elongation, lost protrusions, altered curvature. |
| |
|
| | - New dataset, ATLAS-1: Drug-treated 3D melanoma cells in collagen (control, nocodazole, blebbistatin) to evaluate both performance and interpretability on real data. Available on Hugging Face 🤗 |
| |
|
| | - Higher accuracy than black-box models. |
| |
|
| | - Plug-in head: Works with popular point-cloud backbones. |
| |
|
| |
|
| |
|
| | Try it for yourself: |
| |
|
| | 🌐 Project page + live demo: https://sentinal4d.github.io/PointMIL/ |
| |
|
| | 💾 Code: https://github.com/Sentinal4D/PointMIL/ |
| |
|
| | 🧪 Dataset (ATLAS-1): https://sentinal4d.github.io/ATLAS-1/ |
| |
|
| | If you work on cell morphology, drug discovery, or explainable 3D AI, we’d love your feedback and collaborations. |
| |
|
| |
|
| |
|
| | #ICCV2025 #Sentinal4D #PointMIL #3D #CellBiology #Morphology #ExplainableAI #XAI #ComputerVision #MIL #OpenScience |
