Why does reading biopsy slides take so long?
When a doctor takes a small piece of tissue (a biopsy), it is cut into very thin slices, stained and placed on a glass slide. A pathologist, the doctor who specialises in these slides, studies them under a microscope to say whether there is cancer. It is careful work that takes years of training and a lot of time.
Slides can now be scanned into huge digital images, made of billions of pixels. Artificial intelligence looks like a good helper for reading them. But there was a catch: to learn, these programs needed experts to draw the cancerous areas by hand, slide after slide. That is slow and very expensive.
How did the researchers train the AI without detailed annotation?
The team used what is called weakly supervised learning. Instead of outlined areas, the AI receives just one piece of information per slide: the diagnosis already written in the hospital report (cancer or no cancer). It is then up to the AI to find, among thousands of small image patches, the ones that explain that diagnosis.
The researchers used 44,732 scanned slides from 15,187 patients, without sorting or cleaning the data first. They tested the tool in three situations: prostate cancer, basal cell carcinoma (a common skin cancer) and breast cancer that has spread to the lymph nodes under the arm.
What results did the AI achieve?
To judge how good a test is, researchers use the area under the curve: 1 means a perfect test and 0.5 means guessing. Here it was above 0.98 for all three cancer types.
The authors also simulated a real-life use: the AI sorts the slides and the pathologist only looks at those it flags as suspicious. In that scenario, 65 to 75% of slides could be set aside while keeping 100% sensitivity, meaning no cancer was missed in the data that was tested.
Key figures
- 44,732digitised biopsy slides used, from 15,187 patients
- above 0.98area under the curve for prostate, skin and breast-to-lymph-node cancers
- 65 to 75%of slides that could be set aside with 100% sensitivity
What could this change for patients?
In time, a tool like this could act as a filter: clearly negative cases would be put aside, and pathologists would spend more time on the difficult slides. For patients, that could mean faster results and attention spread more wisely.
Above all, the method shows that such tools can be trained at very large scale with data hospitals already hold, without hundreds of hours of manual annotation. That is what opens the door to similar tools in many other areas of medicine.
What are the limits of this study?
The evaluation used slides that had already been diagnosed. The tool was not tested live with patients to see whether it really improves care. The data came from one large centre, and the results will need to be confirmed in other hospitals, with other scanners and other ways of preparing slides.
Finally, the AI does not replace the pathologist: it is presented as decision support. A sensitivity of 100% in one data set does not guarantee that no cancer will ever be missed in real life.
Waiting for a biopsy result, with Mima
This research AI could one day act as a filter to help pathologists focus on the difficult slides. Your result is still read and signed by a doctor, and nothing changes yet in hospital.
Waiting for a biopsy is long. Mima keeps the date of the sample and of the results appointment, and you can note how you feel. You can also prepare this question: "Who will give me the result, and what could happen next?"
Frequently asked questions
What is a biopsy slide?
It is a thin slice of tissue taken from a patient, stained and placed on glass. A pathologist studies it to see whether cancer cells are present.
What is weakly supervised learning?
It is a way of training an AI with very little guidance: here, only the final diagnosis of each slide, with no hand-drawn areas. The AI has to find the image details that matter by itself.
Can this AI replace the pathologist?
No. The authors present it as decision support that could sort slides. The diagnosis is still made by the doctor.
Is this tool already used in hospitals?
The study is a test on past data. It lays the groundwork for clinical use but does not yet prove a benefit for patients in routine care.
The original study
Campanella G, Hanna MG, Geneslaw L, et al. “Clinical-grade computational pathology using weakly supervised deep learning on whole slide images” Nature Medicine, 15 July 2019. DOI : 10.1038/s41591-019-0508-1
Read the full study in Nature MedicineThis article summarises a scientific study and is not medical advice.