How Many Images Do You Need for a Medical Imaging AI Project?

In shortWith transfer learning, a few hundred well-labelled images per class can give meaningful research results. Below about a hundred per class, report results cautiously and focus on method rather than headline accuracy.

Why the number varies

How much data you need depends on how hard the task is, how different the classes look, image quality, label quality and whether you classify or segment.

What helps with small datasets

• Transfer learning: start from a model pretrained on millions of images.
• Augmentation: realistic flips, rotations and intensity changes.
• Cross-validation: k-fold gives a more reliable estimate than one small test set.
• Simpler models: fewer parameters overfit less.
• Better labels: 300 carefully labelled images beat 3,000 noisy ones.

What doesn’t help

• Splitting slices from the same patient across train and test. It inflates accuracy and reviewers will notice.
• Reporting only accuracy on imbalanced data.
• Augmentations that create unrealistic anatomy.

Practical guidance

• Under 100 images per class: treat as a pilot study.
• A few hundred per class: workable for classification with transfer learning.
• Segmentation: fewer images can work because each image contains many labelled pixels, but annotation quality is critical.

Combine datasets carefully

Public datasets can add data, but different scanners and protocols create domain shift. Report it, and test on data from a source the model hasn’t seen if possible.

Read the full brain tumor MRI overview or the Python cancer detection pipeline.

My imaging services are priced per 100 images — brain MRI and cancer imaging — and I’ll tell you honestly if your dataset is too small for reliable conclusions.