If your family doctor told you that medical schools have been training surgeons using mostly the same methods for the past 200 years, would you believe him or her? Or would you cringe at the thought and ask: “Where are all the technological advancements science fiction writers and the tech industry promised us?”
Lately there has been a lot of buzz about the fourth industrial revolution. We’re happy to tell you, the revolution is in full swing and it will significantly impact the healthcare industry.
Since a large part of diagnostics relies on medical imaging, it makes sense to pay special attention to this field. Groundbreaking systems utilizing AI have already been installed at prominent medical centers. Furthermore, algorithms—most of which are based on IBM’s Watson—are being fed patient records.
Computer vision developers from InData Labs state that the goal of computer vision for healthcare is to reach such a level of sophistication that even mundane devices like smartphones with a camera will become medical diagnostic tools. Medical apps will become a reality and wearables will replace many visits to the doctor’s office.
Computer vision (CV) through machine learning can perform image classification, segmentation, object detection, 3D rendering, and more.
The Perks of Computer Vision in Healthcare
Using a computer instead of a trained specialist in the medical sector could potentially be life-threatening for patients. That is why this technology is only a helping hand for medical staff. Yet, the advantages are undeniable and include an increase in quality, accuracy, and predictability. Saving diagnosis time and early detection of certain diseases are also important, alongside the reduction of costs.
Quality and Replicability
While diagnosis relying on human experience can vary greatly, the accuracy provided by algorithms is superior and easily replicable over similar data sets. Cameras and GPUs are never tired and if the underlying model is executed correctly, they can pick-up details which are easily missed by the naked eye.
Even more, the same algorithm can be used in different hospitals, eliminating travel time and expenses for a highly trained specialist or forcing patients to visit a certain hospital.
Lifesaving Time
Early detection of life-threatening conditions like cancer can mean the difference between life and death. Some forms evolve very rapidly, in a matter of weeks or months. By the time the condition becomes visible, it could be too late for a treatment to be effective. Computer vision systems can be trained to detect even leading indicators like the increased darkness of the skin or irregular contours for moles.
Microsoft’s InnerEye project aims to identify tumors using 3D imaging. The deep learning algorithm can identify the difference between dangerous tumors and benign ones. Philips has another tool, called Illumeo based on adaptive intelligence which is contextually relevant, providing the patient’s previous information. Such systems are usually modular and can be extended and enhanced, even in an Open Source arrangement.
Cost Reduction
Misdiagnosis can cost the medical system thousands of dollars due to the wrong treatments and, more importantly, cost the patient their health due to improper prescriptions. Using CV systems can decrease the rate of misdiagnosis substantially by classifying each condition correctly from the beginning.
During the learning phase, the images are tagged and used to train the system. The same sets of images can be reused for new models, as long as the tags are still relevant. This reduces the time to develop new ways of diagnosing diseases. With just a few changes in the algorithm, the same model can be repurposed from one type of cancer to another in a matter of days.
How It’s Made
Computers don’t understand images in the way we do since this is only a line of 1s and 0s for them, so the only way is to look for patterns. A first approach is to detect the object by identifying borders which are usually of another color, with the intermediary step of region detection. Another way of making sense of an image is segmentation, which means classifying each pixel following a probability distribution which states how reasonable is to assume that it belongs to a certain class. Other techniques are available, including deconvolution.
Potential Drawbacks
Creating excellent CV solutions for the medical industry is not a smooth process. It involves research challenges, data availability and data manipulation problems and other issues, like algorithmic complexity.
The Black Box
All deep learning applications have the disadvantage of no clear explanations for how a specific decision emerged. This hinders the calibration process and makes training the algorithm a matter of trial and error. Even a correct decision can’t be justified, which makes the patients doubt the accuracy of their results.
Also, since treatment relies on the result of the CV, having no trace of the decisional process can result in the wrong medication plan.
Data Availability & Manipulation
As mentioned above, to effectively train a CV model vast amounts of data are necessary—in this case, 2D or 3D images—and more importantly, they should be properly tagged. Variety and slight differences are also required for classification purposes. Unfortunately, such data is not always available. Free, publicly available information is not enough to fine-tune such models, and specific data is most of the times protected by privacy regulations.
To advance the development of these methods, some legal changes might be necessary to use existing and future medical records as training material.
Manipulating data is another challenge since for each application it has to be put in a proprietary format and properly tagged.
Algorithm Improvement
Currently, the acuity of image recognition algorithms is not satisfactory. Also, the automatic learning process needs a lot more improvement. As existing ways of developing CV require expert knowledge, a possible refinement would be to have self-training systems. Future development should include more user-friendly approaches so that the need for large cross-functional research teams is smaller.
