Computational pathology is a rather new term that has been gaining a lot of limelight due to the number of possibilities associated with it. Computational pathology in a very basic sense is the use of mathematical models to make faster and more accurate pathological decisions.
The emergence of deep learning and more importantly the technological advancements that can support the complex models of deep learning are what has made computational pathology a science of today and not the future.
While the number of people at risk of various diseases increases, the number of medical professionals to diagnose these diseases will also have to increase. On the contrary, we already face an acute shortage of Pathologists. This is where technologies like deep learning step in. Deep learning is a part of machine learning that uses data representations to make certain decisions. Once trained these deep learning models can put out results in a few minutes or maybe even lesser depending on the type of sample that is being inferred.
As exciting as this might sound, there is a significant amount of pre-processing that has to be done to data before it can be used by these complex models. These challenges increase two-fold when you’re dealing with un-digitized data. In the context of cytology-based screening, all of the data that exists today is on Glass slides that are in huge archiving libraries of hospitals or laboratories. Apart from the need to digitize significant amount of data, there are numerous operational hurdles before you’ve got your hands on some samples required to train your models. To name a few; ethical committee approvals take time, this can take anywhere between a couple of months to a year. This is because these committee meetings are scheduled biannually in some major hospitals. Next, since digital pathology was ahead of its time for countries like India, there doesn’t exist a lot of digital data.
Once the samples are acquired, they need to be digitized. This can be a very expensive affair since not many digitization platforms or devices exist. Needless to mention, this is time-consuming depending on the quantity of data you would want your models to be trained on.
After this comes to the process of annotation and labelling the digitized data, this needs to be done by experienced pathologists for better specificities and sensitivities. With this comes the most difficult challenge with training data, interobserver variability. A decision to eliminate the discordant data may not always be the right choice, because this means that we might be throwing away data that is challenging to infer.
All in All, Computational Pathology is definitely an answer to the shortage of pathologists and the absence of pathologists in far away areas. Since we’re at the onset of this technology, identifying challenges and solving them quickly will be key for it to flourish. An amalgamation of faster Government and Institutional approvals, co-operative senior pathologists and an understanding of the benefits of this technology prove to be vital.
By Atul Bisht