AUTONOMOUS RESEARCH · MICROSCOPY · SPOT DETECTION

A field of faint signals —
which ones are real spots?

Fluorescence images contain small bright signals against an uneven background. Finding their centers means deciding what is a spot, what is noise, and how that decision changes from one dataset to another.

THE FINDING

Eight evaluations improve detectionacross all nine datasets.

KEY CONCLUSIONS

Aggregate improvement rises from -0.0469 to 0.1158.

The largest F1 gain is in telomeres: 0.6796 → 0.8942. The agent combines detector selection, peak tuning and targeted retraining.

Enter the research workspace
ONE FIELD. EVERY DETECTION IN VIEW.Switch the overlay
Retained DNA-FISH field · F10.923196 matched · 13 unmatched · 3 missed3-pixel matching cutoff · Separate comparison run
DNA-FISHPrediction agreement
img_000.tif
FULL IMAGE · ORIGINAL CONTRAST · PIXEL COORDINATES
CONVENTIONAL RESEARCH

The researcher tunes each detector, one dataset at a time

Choose a method, train it, find peaks, evaluate predictions and revisit the settings when a different kind of image behaves differently.

AGENT RESEARCH

Evaluation feedback becomes the next experiment

The agent changes inference, retrains weak instances and checks the new predictions, carrying each result into the next submission.

WHY IT MATTERS

From a bright pixel to an inspectable detection

The result is a set of spot coordinates, accompanied by per-dataset scores and a record of the changes that produced them.

01 / FROM QUESTION TO EXECUTION

One workspace,
one detection study.

The task is to find spots across nine microscopy datasets.
Follow the recorded changes from the first detector to the final evaluation.

SPOT DETECTION / FROM IMAGES TO EVALUATION
00:00 / 00:32

Rebuilt from recorded changes and scores · Time-compressed animation, not a screen recording

Find fluorescence spots across nine datasets, then use evaluation feedback to refine the detector.

8evaluation submissions
9datasets, including 2D and 3D
72original per-dataset F1 scores
+0.1627absolute aggregate improvement
DETECT. EVALUATE. RETUNE. REPEAT.

The agent connects model training, spot localization and evaluation, then uses the feedback to choose the next change.

02 / WHAT WAS FOUND

Which changes
help find the spots?

F1 measures the balance between finding reference spots and avoiding false detections; higher is better.
Compare all eight submissions within each dataset, then inspect retained predictions from four datasets at the level of individual spots.

01 / THE LARGEST GAIN

Telomere detection improves the most

F1 rises from 0.6796 to 0.8942, a gain of 0.2147. This is the largest first-to-final increase among the nine datasets.

02 / DIFFERENT DATA, DIFFERENT CHOICES

A later submission is not always better for every dataset

DNA-FISH reaches 0.8382 at submission 4 and ends at 0.8358. The per-dataset view preserves these changes, including small declines.

03 / THE FINAL COMPARISON

All nine datasets finish above their starting point

The aggregate improvement rises from -0.0469 to 0.1158. The individual F1 scores show where the gains occur.

DNA-FISH

Eight original F1 scores, in submission order. Select a point to see the recorded change.

Original F1 scores · No best-so-far smoothing

View every original F1 score
Per-dataset F1 · All eight submissions
Download scores ↓Download display data ↓
Turn a detection choice into a testable result.

From choosing a detector to checking its output, every submission leaves a record. The optimization scores and the separate image comparison answer different questions: how the method changed, and what each retained prediction detected.

Explore more cases ↗
About this research
QUESTION, APPROACH AND EVIDENCE

How can a detector improve
across different kinds of images?

The recorded study combines U-Net detection with a Laplacian-of-Gaussian fallback, then revisits peak thresholds, test-time augmentation, training and preprocessing in response to evaluation feedback. All eight submissions and 72 original F1 scores are retained.

What the evidence shows

Optimization result

The confirmed aggregate endpoints are -0.0469 and 0.1158, an absolute increase of 0.1627. Corrected intermediate aggregate scores were not supplied; the evaluation chart therefore uses original per-dataset F1 scores.

Prediction comparison

The first test fields from DNA-FISH, Telomeres, HybISS and MERFISH come from the same separate 20261015 seed-1 run. They are not assigned to optimization submissions. Selection follows filename order, before inspecting scores. The supplied DNA-FISH comparison is unchanged. For the additional fields, stored coordinates are matched with the same Hungarian rule and 3-pixel cutoff; both panels share a linear contrast mapping from the image minimum to its 99.8th percentile. These metrics describe individual images, not whole datasets.

Process animation

The 32-second animation reconstructs the recorded method changes and evaluation results. Diagrams illustrate the method; the numbers are original scores. It is not a screen recording, a new training run or a reconstruction of unavailable intermediate images.

Task s41592-025-02662-x · One optimization run · No claim that a single change alone caused a score difference. Model attribution is omitted.

Evaluation scores ↓Display data ↓Source manifest ↗