Pointmil

From motion to meaning, one frame at a time

Before we can read biology in motion, we need to understand what each frame is telling us. PointMIL, selected as a highlight at ICCV 2025, brings interpretability to static 3D shapes, so every point in a cell can be linked to a model’s decision – a crucial first step towards truly understanding 4D cell dynamics.

At Sentinal4D, we ultimately care about how living cell shapes change over time under different treatments and environments. But every 4D experiment is built from individual 3D snapshots. Each snapshot is a point cloud: a dense constellation of points describing the cell surface at a single moment.

Most point-cloud classifiers today are very good at telling you what something is – a particular cell state, a vessel type, a drug condition – but not why the model reached that conclusion. For drug discovery and mechanistic biology, that is not enough. Teams need to see the evidence: where on the cell surface the model is “looking” and which features drive a prediction.

PointMIL is our answer to that problem for static 3D point clouds:

An inherently interpretable framework for 3D point-cloud classification that improves accuracy while giving point-level explanations as part of the model’s forward pass.

No saliency tricks. No extra explainer network. The same quantities that make the prediction also tell you which parts of the cell shape mattered.

What is PointMIL?

PointMIL treats each 3D shape – a cell, a vessel, or an object – as a bag of points. Instead of collapsing the shape into a single global feature and hoping for the best, it keeps track of how each point contributes to the final class.

Under the hood, PointMIL does two main things:

1. Feature extraction
   A point-cloud backbone (e.g. PointNet, DGCNN, CurveNet, Transformer) turns each point into a feature vector capturing local geometry and context.
   
   

2. Multiple Instance Learning (MIL) pooling
   Rather than pooling all features uniformly, PointMIL uses MIL to learn how to combine point-level evidence into a single prediction. Different pooling modes express different assumptions about how “evidence” accumulates in a static shape.

In all cases, the output is the same pair of objects for a single 3D snapshot:

• Shape-level class probabilities – what the model thinks this cell (or object) is.
  
  

• Point-level importance scores – how much each point contributed to that decision, per class.

Because those scores come directly from the MIL head, interpretation is built into the decision rule rather than bolted on afterwards. These are static shapes – single time points – but they give us a precise, interpretable language for where the model finds morphological evidence. That is the foundation we need before we start reasoning over trajectories and time.

Contextual attention: respecting shape, not noise

One observation from our experiments is that naïve attention can be too sparse on static point clouds. It occasionally latches onto a tiny subset of points and ignores neighbouring regions that clearly matter to a human observer.

To address this, PointMIL introduces contextual attention. After computing attention weights per point, we smooth them over each point’s k-nearest neighbours. This:

• keeps the emphasis on discriminative regions, but
  
  

• respects the fact that morphology is continuous – a bleb, a spine, or an aneurysm bulge is not a single point, it is a local structure.
  
  

The result on static 3D shapes is:

• denser and more biologically plausible importance maps,
  
  

• improved interpretability metrics (AOPCR, NDCG@n) without sacrificing accuracy.
  
  

In other words, PointMIL does not just say “these two points matter”; it highlights coherent regions that align better with how scientists actually think about structure.

Why interpretability matters for cell biology

For us, PointMIL is not just a computer vision result – it is a tool for turning static 3D cell morphology into inspectable evidence.

ATLAS-1: drug-treated 3D cancer cells

To test PointMIL on the kind of data we care about most, we built ATLAS-1 – a public dataset of 1,500 3D melanoma cells imaged with oblique-plane light-sheet microscopy. Each sample is a static 3D snapshot of a cell reconstructed as a point cloud, labelled by treatment:

• No treatment
  
  

• Nocodazole – disrupts microtubules and tends to produce rounder, more compact cells.
  
  

• Blebbistatin – inhibits non-muscle myosin II and often leads to elongated protrusions and altered contractility.
  
  

When we applied PointMIL to these single-frame point clouds:

• For Blebbistatin-treated cells, importance maps concentrated on extended protrusions and blebs, matching expectations for actomyosin disruption.
  
  

• For Nocodazole-treated cells, the model focused on peripheral regions of rounded cells, where microtubule depolymerisation reshapes the cortex.
  
  

• For untreated cells, it highlighted regions lacking these abnormal structures.
  
  

Perturbation analysis on these static shapes showed that removing the points deemed most important for a treated cell often pushed the prediction back towards the “No treatment” class. This suggests that PointMIL is not just separating labels; it is isolating the morphological signatures of drug response in a single frame, in a way that scientists can see, critique, and build on.

ATLAS-1 and the PointMIL code are available for the community, along with an interactive web demo where you can upload your own point clouds and visualise per-point importance directly in the browser.

How PointMIL supports Sentinal4D’s platform

Within SentiCORE, our platform for analysing cell behaviour, PointMIL plays a specific role: it makes static 3D morphologies interpretable.

It helps us – and our partners – to:

• See where the evidence is
  Every prediction on a 3D shape is accompanied by a point-level importance map. Teams can rotate the cell, inspect the highlighted regions, and ask: does this match what we know about this pathway or compound?
  
  

• Connect morphology to mechanism
  Even in a single frame, cell shape carries information about cytoskeletal state, contractility, adhesion, and more. PointMIL makes those signals measurable at the level of individual surface regions, not just as bulk shape descriptors.
  
  

• Support earlier go/no-go decisions
  Because the outputs are visually interpretable, results can be shared across disciplines – cell biologists, chemists, clinicians – and defended in internal reviews. You are not asking colleagues to trust a black box; you are showing them the regions and structures that drive the call.
  
  

• Scale across backbones and datasets
  The MIL head is agnostic to the feature extractor. Whether the upstream model is optimised for speed, robustness, or particular sensors, we can wrap it in PointMIL and retain interpretability without rewriting the whole stack.
  
  

We are not trying to “disrupt” how people do biology. We are trying to make existing workflows clearer, by turning complex 3D data into evidence that teams can see, test, and trust.

What comes next

The ICCV work deliberately focuses on 3D snapshots. That is by design: we first make single frames interpretable, then extend that rigour to full 4D.

Day to day at Sentinal4D, we are extending these ideas towards 4D morphodynamics:

• modelling cell shape over time as sequences of point clouds,
  
  

• integrating PointMIL-style interpretability into temporal models,
  
  

• connecting those dynamic signatures to readouts that matter for our partners: efficacy, heterogeneity, and potential clinical success.
  
  

As a next step, we have developing MorphoSense, which builds on the foundations of PointMIL to classify 4D point clouds of living cell shape changes over time, with the same commitment to interpretability. PointMIL teaches us where the model finds evidence in each frame; MorphoSense will extend that insight across time.

Closing thought

Life happens in motion, but understanding it starts with single frames. With PointMIL, we make those frames legible: point by point, shape by shape, drug by drug.

If you work with 3D or 4D cellular data and want to understand not just what your models predict, but where the biology is hiding on the cell surface, we would be happy to talk.