Cutting the objects out, which is a decision about the data | A crop needs a label, and the label is a decision | What the detector already answers | The classifier, and the metric that is not accuracy | Scoring per class, because the mean is the commonest class | Two things that do not help, measured rather than assumed | The class that is not a picture of anything | What this does not do | Notes on the machine

Before any of that: boxes in a few lines | And here, unlike the nuclei vignette, a score means something | One refusal worth meeting now | Anchors, which a detector has and a U-Net does not | What the loss is doing | Scoring, and the number that is missing on purpose | [email protected], mAP@[.50:.95], and which to believe | Looking at it | What this does not do | Notes on the machine

Before any of that: boxes in a few lines | And here, unlike the nuclei vignette, a score means something | One refusal worth meeting now | Anchors, which a detector has and a U-Net does not | What the loss is doing | Scoring, and the number that is missing on purpose | [email protected], mAP@[.50:.95], and which to believe | Looking at it | What this does not do | Notes on the machine

Before any of that: a mask in four lines | The data | Look at it before training on it | The specification | Train | Compare | Look at what it did | The same thing from a file | A note on older graphics cards

Why a vessel cannot be resized | The data | The configuration | Training | One number hides the finding | clDice beside Dice | The worst case, drawn | Notes

Making something to segment | Splitting by patient, not by scan | The specification | Training | Scores per patient, not per dataset | Looking at a slice | Millilitres, not voxels | Writing a mask back onto its scan | DICOM, which is how the data usually arrives | What this leaves out

One file, one run | Several models at once | Dividing one folder | Augmentation, callbacks, and the rest of a model | Instances instead of one mask | Detection instead of segmentation | Not training at all | What is checked, and when | The same run from code

One file, one run | Several models at once | Dividing one folder | Augmentation, callbacks, and the rest of a model | Overlapping classes instead of one per voxel | When the overlap was drawn separately: classes_from | Instances instead of one mask | Detection instead of segmentation | One label instead of a mask | When the label is a set: labels_from | Predicting on images you have no labels for | Not training at all | What is checked, and when | The same run from code

Why a circumference is not a score | The data | Splitting by examination, not by image | The specification | Training | From a mask to a millimetre | What the annotation costs before the model does | The Dice cannot see this error | What this does not settle | Notes on the machine

Before any of that: a mask in four lines | The data | Look at it before training on it | The specification | Train | Compare | Look at what it did | The same thing from a file | A note on older graphics cards

Why a vessel cannot be resized | The data | The configuration | Training | One number hides the finding | clDice beside Dice | The worst case, drawn | Two of those choices, measured afterwards | The loss is a decision, and this configuration already had the right one | A surface loss is the one that breaks the lines, and it gets worse with the dose | Tiling against resizing, where this document's first section is too strong | The rig, which is worth more than the ranking | What none of this says | Notes

Making something to segment | Splitting by patient, not by scan | The specification | Training | Scores per patient, not per dataset | Looking at a slice | Millilitres, not voxels | Writing a mask back onto its scan | DICOM, which is how the data usually arrives | What this leaves out

Before any of that: objects in four lines | The semantic sibling, side by side | The data | The specification, with the borrowed model beside a new one | Train | Compare | What this does not settle, measured rather than guessed | Which objects, not how many pixels | A decoding changed without retraining | What it was for: a count and a size distribution | And what shape, which needs no scale at all | The same thing from a file | A note on older graphics cards