Do nitra živé buňky: Mikroskopie a počítačové vidění - Filip Šroubek (UTIA) [seminář MPN 30.9.2026]

Scientific concepts, discoveries, and nature/biological phenomena

1) Problem framing: tracking in microscopy vs. tracking in video

Computer vision “tracking” / “monitoring” generally means following identities of moving objects over time.

Compared to street-camera person tracking, microscopy tracking adds several difficulties:

• Low resolution: objects may be only ~10–30 pixels wide.
• Low signal-to-noise ratio (SNR): high noise (analogous to “night / black-and-white” conditions).
• Limited frame rate: few tens of frames per second, requiring tracking under temporal sparsity.
• Shape changes: cells/vesicles can change morphology, unlike people who typically change shape only minimally.

---

2) Cellular transport phenomena: endocytosis and exocytosis

• Endocytosis: the cell uptakes cargo (e.g., nutrients, and even viruses) by forming vesicles from the membrane.
  • A mentioned mechanism includes clathrin, which forms a scaffold that drives vesicle formation.
• Exocytosis: the cell releases vesicle cargo outside the cell by vesicle fusion with the membrane.

Biological objects of interest:

• Vesicles of about ~100 nm diameter, often transparent/invisible in standard optical microscopy.

---

3) Fluorescence microscopy principles (molecular labeling)

Fluorescence microscopy uses fluorescent markers that bind selectively to target proteins. When proteins are labeled, cellular events become visible as glowing signals.

Example labeling strategies mentioned:

• Label vesicle-forming proteins involved in exocytosis.
• Label clathrin scaffolding associated with endocytosis.

---

4) Limits of classical optical microscopy and the diffraction limit

Classical fluorescence microscopy provides insufficient spatial detail:

• Approximate reported resolution: ~200–250 nm per pixel.
• Vesicles appear as blobs or low-detail shapes, making fine morphological analysis difficult.

---

5) Super-resolution microscopy (SIM) to surpass the diffraction limit

SIM (Structured Illumination Microscopy) aims to go beyond the diffraction limit.

How it works:

• Illuminate the sample using structured light (interference “stripes”).
• Use reconstruction to recover higher-resolution information.
• Employ Fourier-domain reasoning: shifting high-frequency information into measurable regions.

Lecture description included:

• The sample is illuminated 9 times with different structured-light patterns (different orientations/frequencies).
• Reconstruction is described as doubling resolution (as claimed in the talk).

Reported approximate achieved resolution:

• ~100 nm resolution
• ~30 nm per pixel (in the example scale)

TIRF/SIM combination:

• TIRF (Total Internal Reflection Fluorescence) excites fluorophores only in a thin region ~100 nm near the membrane.
• TIRF + SIM enables high-detail vesicle morphology while restricting excitation to the membrane region.
• Vesicles can appear as:
  • rings in TIRF-like geometry, and
  • later as blobs as morphology changes.

---

6) “Shape to Fate”: morphological change predicts event success

A central biological/computational idea is that vesicle morphology and its time-dependent shape changes determine whether an exchange is:

• productive/successful, or
• abortive/failed.

This is reflected in the paper title concept: “shape to fate” (shape → fate).

---

7) Multi-object tracking (MOT) and pipeline for vesicle-event analysis

The talk outlines a pipeline with three main steps:

(1) Acquisition

• Imaging using specialized microscopes, notably TIRF + SIM.

(2) Reconstruction

• Super-resolution requires computational reconstruction using Fourier transforms (an inverse problem).
• Practical issue: sequential acquisition can cause artifacts if cells/vesicles move during the multiple illumination frames.

(3) Detection + linking

Object detection (vital for vesicle instances):

• Uses neural networks, especially instance segmentation / object-level detection concepts.
• Manual labeling is hard because microscopy labels are expensive to obtain.
• Approach described:
  • Train using synthetic data from a simulator/generator (since physics/modeling is more tractable for the microscope and reconstruction).
  • The network predicts distance maps.
  • Instances are extracted as local maxima.
  • Synthetic modeling includes vesicle morphology dynamics (e.g., ring → blob behavior).

Linking detections into tracks:

• Formulated as multi-object tracking (MOT).
• Uses a global optimal linking method as a graph problem:
  • detections are nodes across frames,
  • edges represent plausible correspondences based on:
    • spatiotemporal proximity, and
    • (importantly) shape/morphology changes.
• Solved as a minimum flow / integer linear programming problem (per the talk’s description).

---

8) Machine learning methodology trends mentioned

• Instance segmentation: useful for countable objects; can provide masks in addition to detections.
• Supervised learning limitations in microscopy:
  • labeled microscopy data is scarce and difficult to produce.
• Synthetic training data:
  • used when accurate forward models/simulation are feasible.
• Foundation models / Transformer-based approaches:
  • mentioned as a current trend (e.g., clustering image patch features into objects),
  • with the caveat that task-specific supervised training can outperform generic foundation models.

---

9) Downstream biological/statistical interpretation from tracks

After tracking vesicle events across many cells (thousands of trajectories), biologists can test hypotheses such as:

• If an exocytic event occurs at a location/time, is there a higher probability of nearby/succeeding endocytic vesicles shortly after?

The computer vision output supports:

• trajectory-based statistics, and
• spatial/temporal correlation analyses.

---

10) Additional practical tools and data availability

The talk mentions providing online tools for biologists, including:

• Shape to Fate tool (previously described).
• A segmentation tool (“M*Sphereosek” / likely a misspoken tool name), using multiple segmentation methods.

It also emphasizes:

• creating new public datasets, and
• uploading them to platforms like Zenodo for training.

---

Researchers or sources featured (named people/institutions)

• Filip Šroubek (speaker; Institute of Information Theory and Automation)
• Adam Harm (doctoral student / main author)
• Zuzana Karlecová (biologist, Masaryk University)
• Masaryk University
• University of Cambridge (England)
• UTIA (Institute of Information Theory and Automation; referenced in context)
• GATHER / “Gachter agency” (sponsor mentioned; exact agency name unclear)
• Biocev (Paris/Prague mentioned; exact context unclear)
• Institut Curie (Paris) / Curie Institute in Paris
• MTF/Fyzika / Math Faculty of Physics (offered courses mentioned; not a person)
• Nobel Prize 2014 (for super-resolution fluorescence methods; specific recipients not named in subtitles)
• PyTorch (ML framework mentioned)
• TensorFlow (mentioned and contrasted; not used in their workflow)
• Zenodo (public data hosting platform mentioned)
• vibe coding (AI coding tool mentioned; exact tool name unclear)

(No additional named scientists were clearly specified beyond the above; several references are to methods/awards rather than individual researchers.)