engineering careers  AI learns to detect tumours in microscopy images
engineering careers  AI learns to detect tumours in microscopy images

A team Japanese researchers have been able to spot different cancer within microsoptic images using deep-learning AI techniques. The new technique could lead to better cancer treatments.

The technique used ‘phase-contrast microscopy images’ – a technique that reveals cellular structures that are not visible with a simple microscope. The algortium was able to spot different cancer cells with 96% accuracy.

The team, from Osaka University, used common type of deep learning (convolutional neural network or CNN) to take the images and apply a set of connected filters and math functions in order to train the algorithm to spot specific features within cells.

Like the human brain – a CNN can both focus on both fine details such as edges within the image, while also being ‘aware’ of complex features across the whole image.

How to train an AI to spot tumours

Training AI to perform medical diganoistics is not new. However, the teams technique is unique because it could distinguish the fine details in tumours.

These fine details can often mean a tumour is resistant to certain types of treatments. Being able to clasify the type of tumour before treatment allows doctors to tailor a treatment towards a specific patient.

For example, if a tumour is going to be resistant to radiotherapy then it is a waste of time to treat a patient with radiation.

Up to now tumour classification have been mostly performed by visual inspection which takes u[ valuable time and opens the process to human error.

How the new technique works

The new technique instead allows the CNN to break up cells into one of five catergories.

  • Untreated (control)
  • X-ray-resistant mouse tumours (NR-S1 type)
  • Carbon-ion beam-resistant mouse tumours (NR-S1 type)
  • Untreated human cervical tumours (ME-180 type)
  • X-ray-resistant human cervical tumours (ME-180 type)

The team was able to train the CNN from a database of 8,000 phase-contrast microscopic images. Then validated the resulting algorithm using another 2,000 strong data set.

The result was a algorithm that can recognise mouse tumours (NR-S1) then figure out if the cells had developed resistances against commonly treatment therapies such as X-ray or carbon-ion beams with 96% accuracy.

Next steps

While the results are promising the CNN now needs to be trained to better recognise human cells and spot other types of cancer cells. The teams ulimate goal will be to establish a universal system that can spot a whole range of tumours in human cells.


Published in Cancer Res. 10.1158/0008-5472.CAN-18-0653.