Radiologists can now see images as they currently are, and if necessary, request for an opinion from artificial intelligence.
There are various challenges when screening mammography. For instance, some women develop breast cancer but are required to undergo additional screening tests, such as MRI and ultrasounds. This makes diagnosis costly which also adds stress to patients.
Deploying AI in Radiology
A team of researchers at the NYU School of Medicine introduced AI-powered effort in an attempt to tackle the challenges faced in screening mammography. They were led by Dr. Krzystof J. Geras who is an assistant professor in the department of radiology in the institution.
According to him, they intended to decrease the number of additional imaging using AI. He notes that radiologists miss a small fraction of cancer cases during screening mammography examinations. They are hopeful that their AI tool will help them discover such cases which will help in saving lives.
ResNet-22 is the name of the proposed technology. It involves a type of deep convolution network and works by learning from a huge number of image/label pairs. According to Geras, they have trained the network by presenting it with more than 800,000 samples consisting of correct diagnosis outcome several times. It took Dr. Geras and his team approximately 3 weeks to train their program using a very powerful computer that has a graphical processing unit.
AI to Help Radiologists
The NYU School of medicines expects that the AI technology will become an assistant to radiologists.
Geras says radiologists will now see the images as they see them, and if necessary, seek the opinion of the AI. The AI is capable of giving radiologists a predicted probability if the patient has a problem and point the parts that appear most suspicious if there are some.
The team is hopeful that their technology will boost the confidence among radiologists and lower the number of patients who are requested to undergo additional tests.
Dr. Geras explained that their AI has not been deployed in their hospital and that the results that were in a paper they published come from a retrospective reader study.
The Reader Study
The reader study involved radiologists who were requested to air their predictions of probabilities as to whether the patients had breast cancer using the results of mammography screening images. They requested the AI to do just the same, Geras said. After the team averaged the predictions of radiologists and the AI, the team found something interesting. The average was more accurate than either of the two separate predictions. According to Dr. Geras, AI and radiologists were using different features of the data.
A Pilot Study
Dr. Geras stated that their technology can be easily integrated with the real clinical pipeline and that they are currently considering a pilot study at NYU Langone. This will be an attempt to validate it in a clinical setting.
The greatest achievement that the team from the NYU School of Medicine made was enhancing the accuracy of AUC (a measure of accuracy) of radiologists from approximately 0.8 to 0.9 with the help of AI.
“A random predictor achieves an AUC of 0.5 and a perfect predictor achieves an AUC of 1.0,” Geras explained. They were able to achieve such strong results with the help of the large size of data they set when training their neural network. It is impractical for a radiologist to compare such many images in the final diagnosis.
Geras explained that the neural network learnt tirelessly for 3 weeks, and the team hopes to keep accumulating more data. This will help to improve the results.
Geras AI Expert Advice
Healthcare providers working with AI and related technologies can learn some advice from Dr. Geras.
Extensive testing of AI technology before using them in clinical practice is very critical. Although AI technologies are quite promising, they were using deep neural networks that are very sensitive to the changes of data distribution. Currently, they can’t offer any guarantees that the AI will be accurate to different populations and if the images are captured using other imaging equipment.
Geras also notes that there are plenty of companies that are aggressively marketing AI technology as a way of solving the many challenges in healthcare. He insisted that it is important to consider the safety aspect of AI technologies before full implementation.
Dr. Geras predicts that achieving clinical value and routine use will take some years. He advises that AI should not be a monolithic thing.
Finally, Dr. Geras notes that different AI solutions may give completely different results in their performance and robustness. They can also differ a lot in the levels of their predictions. This implies that a lot of research is necessary in neural networks and medical imaging for the technology to realize its full potential, he concluded.