Blog Viewer

Aiding the imaging with Machine learning in medical sciences

By Sayan Basak posted 07-05-2020 08:57


The realms of histopathology still face numerous challenges owing to the difficulty in interpreting the slides by the pathologists. Furthermore, the techniques like microscopic studies do not allow the detection of specific genomic driven mutations, especially in untraced locations (such as subcellular structures of the cells). If we consider using biomarkers, even then, we need specialized instruments or specific complementary setups to integrate the use of biomarkers- which often is not accessible ( in a country like India).  The combination of these problems eventually led to the growth in the use of machine learning to use label-free cell-based diagnostics that are not only accessible ( from anywhere), but also can precise results which couldn’t have been possible with humans alone. 

The latest technology often cited as the ‘High-throughput imaging’ allows the science to resolve cell morphology by microscopy, which often is employed to investigate moieties that change the morphological conditions when applied with a stimulus. A human eye has an inherent error property; even the best analyst is deployed to study the morphological pattern. And, as we say that a slight error in medical sciences may lead to a disaster; the incorporation of machine learning ultimately makes the error percentage obsolete, thus tuning the process to make more precise and robust. Moreover, machine learning provides several supplementary data which can be used as a primary or as a secondary analysis tool to predict those chemicals targeting a different pathway or biological process.

How is this achieved?

Well, in 2018, a research team headed by Dr. Gustin developed a Bayesian matrix factorization method that reported to predict the activity of various unrelated protein targets using the glucocorticoid HTI assay. Currently, the technique relies on the single HTI screen, as the high throughput imaging assays is still in its nascent phases. However, the method requires a library to train itself the probable compounds which it needs to study, which might be claimed as the fundamental disadvantage which the technique has. So, if you miss teaching the algorithm a part of the library, it won’t detect the morphological changes associated with that portion of the content is supposed to circumscribe. We hope that multiple screening HTI algorithms shall be developed sooner to reduce the time required to predict the outputs derived from the fetched inputs. 

The unstained white blood cells were recently sorted using the machine learning algorithms reinforced on flow cyclometry using the label-free technique.  The machine learning techniques like AdaBoost, Gradient Boosting, K-Nearest Neighbors, RF, and SVM amalgamates the statistical prediction models which the sensitivity and the imaging ability derived from the microscope. 

We predict that deep learning may be used to enhance the process further, as deep learning has been proven to work efficiently in models that use cellular images as they supplement the process of learning. Google has already benchmarked convolutional neural networks to predict the Gleason scoring of prostate cancer cases. The initial studies have predicted that the machine learning-driven model provides a diagnostic accuracy of 0.70 as compared to the mean accuracy of 0.61 ( traditional analysis).