Frequently Asked Questions

Images (required)

The microscopy or sample images you want to analyse. Standard formats such as TIFF, PNG, and JPEG are supported. Every other file type is optional and supplementary: images are the only thing you must upload to start working.

Masks (optional)

Segmentation masks that outline the cells or regions of interest in your images. Each mask is a labelled image where every cell is represented by a unique integer ID, and background pixels are zero. Uploading existing masks lets you skip the automatic segmentation step and jump straight to classification or analysis, or to continue refining masks you created in a previous session. You can upload the Cellpose .npy format or a tiff where each cell is a different integer value.

Cellpose model (optional)

A trained segmentation model as a .pt or .pth file, either fine-tuned in Mycol and downloaded earlier or trained elsewhere. It then replaces the default model everywhere segmentation runs. Without one, Mycol falls back to its built-in pre-trained model, which is usually a reasonable starting point.

DenseNet model (optional)

A trained cell classification model, also .pt or .pth. Unlike segmentation there is no default classifier, so automatic classification stays unavailable until you upload one or train one. Mycol tells the two model types apart from their contents, so both go in the same uploader and neither needs special naming.

Cellpose inference hyperparameters CSV (optional)

A small two-column parameter,value file listing the Cellpose inference settings - diameter, cell-probability threshold, flow threshold, minimum size, and number of iterations. These significantly affect segmentation, so uploading the cellpose_inference_hyperparameters.csv (found in a downloaded model or a saved-session zip) automatically applies those settings on the Annotation page, letting you reproduce a previous segmentation without re-entering each value by hand.

Saved Session (optional)

A zip archive produced by the Session Restore download option. Uploading it reloads a previous Mycol session in full: all images, masks, class labels, trained models, and settings are restored exactly as they were when the session was saved. This is the recommended way to pause and resume long annotation or training projects, or to share a complete working state with a collaborator.

Practical notes

• Masks are paired to images by filename. A mask is named after its image plus a suffix, _masks by default, so plate1.tif pairs with plate1_masks.tif. The suffix can be changed next to the uploader. A mask with no matching image is skipped and reported.
• Images are resized to 512 x 512 by default, to keep the app responsive on large images. Measurements are unaffected: Mycol converts every size back, so areas and lengths always refer to your original image. Untick the box beside the uploader to keep full resolution.
• The limit is 512 MB per upload. Larger sets are fine, just add them in batches; later uploads extend your session rather than replacing it.
• You do not need your own data to start. Use demo data loads a small example set so you can walk through the whole workflow first.

Cellpose is a deep-learning model purpose-built for biological cell and nucleus segmentation. Rather than predicting a simple foreground/background mask, it learns to predict gradient flows - vectors that point from every pixel toward the center of the nearest cell. A simple post-processing step then follows those gradients to group pixels into individual cell instances, making it remarkably robust to cells of different sizes, shapes, and imaging conditions.

Cellpose is pre-trained on a large, diverse collection of microscopy images and generalises well out of the box. When your images differ significantly from the training data (unusual stains, atypical morphology, low magnification), fine-tuning on a small set of your own annotated images - which Mycol supports directly - can dramatically improve results.

Versions used in Mycol

• Cellpose 2 introduced a human-in-the-loop training workflow, allowing users to correct segmentation errors and immediately improve the model with very few additional annotations. Read the Cellpose 2 paper (Nature Methods, 2022) →
• Cellpose 3 added image restoration capabilities (denoising, deblurring, and upsampling) so that segmentation quality is maintained even on noisy or low-resolution data. Read the Cellpose 3 paper (Nature Methods, 2024) →

SAM2 (Segment Anything Model 2) is a general-purpose segmentation model developed by Meta AI. Unlike Cellpose, which was trained specifically on microscopy images, SAM2 was trained on an enormous and diverse collection of natural images and videos. Its strength is promptable segmentation: given a point, bounding box, or rough outline supplied by the user, SAM2 predicts a precise mask for the object at that location.

In Mycol, SAM2 is used as an interactive segmentation tool. Instead of running fully automatically across the whole image, you highlight a cell with a box and SAM2 will generate a high-quality mask for that cell. This makes it especially useful for:

• Segmenting unusual cell types or objects where Cellpose may struggle.
• Quickly annotating small numbers of cells for a training set without running a full automatic pass.

Because SAM2 relies on your prompt rather than learning from your specific image type, it does not need to be fine-tuned - it works out of the box. However, for fully automatic, hands-off segmentation of large image sets, a trained Cellpose model will generally be faster and more consistent.

DenseNet (Densely Connected Convolutional Network) is a convolutional neural network architecture designed for image classification. Its key idea is that every layer is directly connected to every other layer within a dense block: the input to each layer is the concatenation of the feature maps from all preceding layers, not just the previous one.

This dense connectivity has two practical benefits:

• Feature reuse - early low-level features (edges, textures) remain accessible to later layers without having to be re-learned, so the network extracts richer representations with fewer parameters.
• Gradient flow - during training, gradients can travel directly from the loss back to the earliest layers, which prevents the “vanishing gradient” problem and makes training more stable, especially on small datasets.

Mycol uses DenseNet-121, applied at the individual cell level: each segmented cell is cropped out as its own small patch and classified into one of your user-defined classes.

Training never starts from scratch. Mycol begins from a backbone pre-trained on ImageNet and replaces only its final layer with one sized to your classes, so the low-level features above are already learned before it sees a single cell of yours. That is why a useful classifier needs tens of labelled cells per class rather than thousands. You can also skip training entirely and upload a checkpoint trained earlier.

A trained model is not a program that understands cells. It is a statistical mapping, fitted to the exact examples you gave it, from patterns of pixels to an output. It has no concept of what a cell is. Nearly every disappointing result traces back to the same root cause: the images or the labels used for training did not match what the model was later asked to do. The sections below cover the ways that happens, and how to avoid them.

A model learns to be good at exactly what you trained it on

When you look at a dim, oddly lit image and still see the cells instantly, you are using knowledge the model does not have. All the model learned was which pixel patterns earned a reward on your training images. Change those patterns - even in ways you would not consciously notice - and it is being asked to extrapolate into territory it has never seen. It has no fallback of “well, obviously that is a cell.”

This is why a model that worked perfectly last week can fail on today’s plate while the task looks completely unchanged to you. The task is unchanged; the input statistics moved.

What counts as a “different” image

Far more than most people expect. Any of the following can be enough to degrade a model, even when the biology and the question are identical:

• Lighting and exposure - a different lamp setting, a change in illumination angle, ambient room light, or automatic white balance on the camera.
• A different microscope, camera, or objective - even two instruments of the same model can differ enough to matter.
• Magnification and pixel size - a cell 30 px across is a different object to the model than the same cell at 90 px. For Cellpose this also means the diameter setting used at inference has to match what the model was trained for.
• Focus, motion blur, and noise levels.
• Background - a different agar or medium colour, a different plate or slide, condensation, scratches, or a clean background where training images had a textured one.
• Stain, channel, or contrast method.
• The sample itself - a new strain, a different growth stage, higher density, or colonies touching where training images had isolated ones.
• File handling - JPEG compression artefacts, 16-bit versus 8-bit, resizing or downscaling, or a different export pipeline from the acquisition software.

Augmentation helps, but it cannot invent data

During training, Mycol shows the model randomly altered copies of your examples. Cell patches are mirrored, rotated by up to 15 degrees, nudged in brightness and contrast, and slightly blurred; Cellpose applies its own equivalent transformations internally. This is called augmentation, and it buys some tolerance for free: you do not need to worry about which way round your cells happen to lie, or about small exposure and focus differences between images.

What augmentation cannot do is invent data. It only perturbs the examples you supplied, so it will never conjure up a lighting rig, a magnification, or a morphology you never imaged. Genuine diversity in the training set is what buys generalisation; augmentation just smooths the edges of it.

When training Cellpose, outline every cell in the image

This is the single most common annotation mistake, and it is easy to make because nothing warns you about it. Cellpose trains on the whole image. Anything not inside an outline is not treated as “unknown” or “skipped” - it is an explicit example of background. A cell you did not get around to outlining is therefore active training data teaching the model that things which look like that are not cells.

In practice this means:

• Skipping the faint, small, clumped, or unusual cells teaches the model to ignore exactly those cells - the hard cases you most wanted help with. The easy ones it could probably already find.
• A half-annotated dense image is worse than no image at all: every cell left unoutlined contributes a wrong signal, and dense images contain many of them.
• Be consistent about cells cut off at the image border. Decide a rule - always outline partial cells, or always leave them - and apply it to every image. Doing it both ways teaches the model that such objects are sometimes cells and sometimes not, and it will hedge accordingly.

Rule of thumb: ten fully annotated images beat fifty partly annotated ones. If an image is too dense to finish, crop a region you can annotate completely and train on that instead. Mycol’s min cells per image setting (default 5) drops near-empty images from training, but do not rely on it to catch partial annotations - an image with twenty of its forty cells outlined passes that filter easily.

Classification labels behave differently - partial is fine, inconsistent is not

DenseNet classifies individual cell patches, so a cell you never assigned a class to is simply left out of the training set. It is not taught as a wrong answer. Partial labelling is therefore safe for classification, unlike segmentation, and you can label as many or as few cells as you have patience for.

What does hurt is inconsistency. If you call borderline cells one thing at the start of a session and the opposite thing two hours later, or two people apply subtly different criteria, the model receives contradictory examples of the same appearance and learns the boundary between your classes as noise. Write down your criteria for the ambiguous cases before you start, and treat “I cannot tell” as a legitimate answer - leaving a cell unlabelled is much better than guessing.

Watch out for imbalanced classes

If 95% of your labelled cells are “normal” and 5% are “abnormal”, a model that answers “normal” for every single cell scores 95% accuracy while being completely useless. A headline accuracy figure hides this entirely.

After training, read the confusion matrix rather than the accuracy figure. Each row shows what actually happened to the cells of one class, so a rare class the model never predicts shows up immediately as a row whose counts have all drifted into the wrong column. The macro-averaged precision, recall, and F1 that Mycol reports alongside it also help, because they weight every class equally rather than by how many cells it has. Aim for a roughly similar number of labelled cells per class; if the interesting class genuinely is rare, deliberately go and find more examples of it to label, and judge the model on how it handles that class.

Good numbers on your own data are not proof that it works

During training, part of your data is held back as a validation set and the reported scores are measured on it. That is a real check that the model has not simply memorised its training examples - but it is a weak one, because the held-back data came from the same session, the same plates, and the same imaging setup as everything else. It tells you the model generalises to more of the same. It does not tell you it will generalise to next month’s plates.

The only test that counts is running the model on a fresh batch of images it has never seen - ideally acquired on a different day, ideally by a different person - and looking at the output yourself. Set a few such images aside at the very beginning and never train on them.

Beware repeated tuning, too. If you re-train several times with different settings and keep whichever version scored best, you have quietly started fitting to the validation set as well. The more times you consult that number to make a decision, the less it means.

More images only help if they are different images

Two hundred fields from one plate on one afternoon are, as far as the model is concerned, close to a single image repeated two hundred times. They mostly make training slower. Twenty images spanning three lighting conditions, two magnifications, and a range of cell densities teach far more.

When choosing what to annotate next, the most informative images are the ones your current model gets wrong. Run it, find the failures, correct those, and add them - the annotate → train → correct → repeat cycle is efficient precisely because it keeps steering your training set towards what the model still lacks.

You do not need a machine learning background to train useful models in Mycol. The guidelines below cover the most impactful things you can do to get good results.

Start with the pre-trained Cellpose model

Before training any segmentation model, run the default Cellpose model on a few images. Often the out-of-the-box result is already good enough, or needs only minor corrections. Training is worth the effort only when automatic results are consistently poor.

Classification has no equivalent starting point: your classes are whatever you define them to be, so no general-purpose classifier can exist, and the automatic classification buttons stay disabled until you train a DenseNet model or upload one. The question is never whether to train, only how many cells to label first.

Data quality beats data quantity

A small set of carefully checked annotations is almost always better than a large set of sloppy ones. For Cellpose, 10–30 well-corrected images are usually enough to see a meaningful improvement. For DenseNet classification, aim for at least 50–100 annotated cells per class, and ensure the classes are balanced (similar numbers per class).

Cover the diversity in your data

Training images should represent the full range of conditions you will encounter: different brightness levels, focal planes, cell densities, and edge cases. A model trained only on the “easy” images will struggle with the harder ones. This matters more than almost anything else - see why models fail on images that look obvious to you.

Watch the loss curves

After training, Mycol shows a plot of how the model’s error (the “loss”) decreased over each training epoch. A healthy training run shows both the training loss and the validation loss falling steadily and then levelling off close together. If the validation loss starts rising while the training loss keeps falling, the model is overfitting - it is memorising your training data rather than learning general patterns. In that case, stop training earlier (fewer epochs) or add more diverse training images.

Iterate in small steps

Train a model, test it on new images, correct the errors, add those corrected images back to your training set, and re-train. This cycle of annotate → train → correct → repeat is the most efficient way to improve a model without needing thousands of images up front.

When Mycol measures a cell - its area, perimeter, diameter, and so on - the raw number is always in pixels, because that is what the image is made of. A pixel has no inherent physical size; it depends entirely on your microscope’s magnification and camera settings.

The pixel-to-distance conversion lets you express measurements in real-world units by supplying a single number: how much physical distance one pixel represents in your images (the pixel size or image scale, often found in your microscope software or image metadata). Turn it on with the Convert to distance option on the Compare Phenotypes page, then enter the pixel size and its unit. The unit is only a label, so any text works; it defaults to µm and follows through to the plot axes and the exported table.

For example:

• If 1 pixel = 0.5 µm and a cell has an area of 400 px², its physical area is 400 × 0.5² = 100 µm².
• Linear measurements (perimeter, diameter) scale by the factor directly: 40 px × 0.5 = 20 µm.

Without a conversion factor, Mycol reports sizes in pixels. Either way this touches only the descriptors that carry a unit: area, perimeter, the two axis lengths and the Feret diameter. Ratios such as circularity, solidity and eccentricity are dimensionless and never converted.

Pixels are fine for comparing cells within one experiment, but cross-experiment or publication comparisons need physical units. Enter the pixel size of your original images: if resize-on-upload was left on, Mycol has already converted every measurement back to original-image pixels.

For every segmented cell, Mycol computes a set of descriptors from the mask region: eight describing its shape and three describing its colour. All of them appear in the downloadable cell metrics table and can be plotted by class on the Compare Phenotypes page. Shape descriptors are computed using skimage.measure.regionprops from the scikit-image library.

Only the size descriptors carry a unit: area, perimeter, the two axis lengths and the Feret diameter are in pixels unless a pixel-to-distance conversion is supplied. The remaining descriptors are ratios or fractions and are dimensionless, so they read the same whatever units you work in.