Skin cancer detection: A Heidelberg algorithm is better than most doctors

Last week, Heidelberg researchers published the data from a study in which they found that a software outnumbers dermatologists in the diagnosis of melanoma. Both the specialists of different German ...

Last week, Heidelberg researchers published the data from a study in which they found that a software outnumbers dermatologists in the diagnosis of melanoma.

Both the medical specialists from various German university hospitals and the software developed by the Heidelbergers evaluated 100 images according to whether it was a harmless birthmark or black skin cancer. In the end, the software was usually more accurate than human diagnostics: only seven of the 157 university dermatologists scored better than the algorithm, 14 got equally good results, and 136 had worse diagnostic results.

Now, just a few weeks ago, the University of Heidelberg announced an alleged world sensation in the field of breast cancer diagnostics, which subsequently developed into a public relations disaster. Also on the skin cancer software, the hospital is involved in a community facility with the German Cancer Research Center and the German Cancer Aid.

Getty Images

Diagnosis at the dermatologist

In the case of the skin cancer formula, however, the all-clear can be given for the time being: The data do not seem exaggerated; Rather, they are just another confirmation of the long-known. Good software is often superior to people in pattern recognition of skin cancer.

As early as 2013, there were publications on how to detect melanoma using smartphone apps, although the early software still underwent many misdiagnoses. In 2017, the breakthrough and one of the most significant lessons in the field of deep learning, ie the use of artificial neural networks, has been achieved: An algorithm called "GoogleNet Inception v3 CNN" was trained with 1.28 million non-medical images, followed by just under 130,000 photos of 2032 skin diseases, which in turn have been compared with biopsies. Twenty dermatologists at Stanford University opposed the system. Software beat every single one.

The Heidelberg researchers also used a CNN algorithm. The abbreviation stands for Convolutional Neural Network, which is a specific architecture of artificial neural networks.

Most skin problems are diagnosed by non-dermatologists - such as general practitioners - and the error rate is reported as up to 50 percent. Melanomas have good chances of recovery if they are detected early and have no metastases in lymph nodes or elsewhere in the body. Thereafter, however, the chances of recovery are bad.

Software can not do everything that dermatologists have in store for diagnostics. Nonetheless, in addition to radiology and pathology, dermatology will become one of the core applications of machine learning in medicine.

You like web world topics? Then subscribe to posts like this one. The newsletter start menu is free and ends up every Monday afternoon in your mailbox.

Strange digital world: talking via smartphone

The other day I was waiting for someone and watched the conversation between an apparently homeless person and a passerby not far from the Charité Campus in Berlin-Mitte. The pedestrian wanted to give the crouching man money, but first talk to him. This was difficult, because the homeless did not understand German. Using Google Translate on the passers-by's cell phone, the two tried to communicate.

When the language was finally identified, the men had a long talk about the app - it seemed fun to them. After about ten minutes, the man left, left a few coins to his interlocutor. He looked after him with a smile.

App of the Week: " Photographs - Puzzle Stories "
tested by Tobias Kirchner

EightyEight Games

"Photographs - Puzzle Stories" is a creative puzzle game that tells interesting and emotional stories. Various photo motifs unlock riddles, which in turn reveal new chapters of a story if successful. For example, it is about the past of a building, which people lived or worked there, and what happened there. The player can always make decisions and take different paths. Often the stories have a melancholy and serious background and are anything but banal.

For $ 4.09 (Android) or $ 4.49 (iOS), from EightyEight Games, with no in-app purchases.

Foreign Link: Three tips from other media

  • Digital fasting and how the platforms respond to it (from 8:15 minutes on, radio report)
    We allegedly touch our mobile phone 2500 times a day. How does digital detox work and does it bring something?
  • How Can We Be Sure Artificial Intelligence Is Safe For Medical Use? (English, 6:49 minutes playing time)
    In the future, more and more medical diagnoses of machines could be made. But how is it guaranteed that the algorithms work properly? The US radio station NPR has researched.
  • Digitization needs to be learned (two minutes of reading)
    The FAZ reports on a study in which 74 percent of the respondents stated that their work had already changed through digitization. Many miss training opportunities in the digital subject area.

I wish you a pleasant week.

Martin U. Müller

ref: spiegel