Machine Learning’s Transformation of Diagnostics and Treatment

Why is this relevant to us? Because no matter what kind of data you collect, it needs to be as distinct as possible and be accessible in bulk. Now, let’s say you manage to collect varying data of medical images while ensuring both quality and quantity, one cannot simply feed it into a program and expect it to function autonomously. Computers, unlike humans, can’t “see” things. Thus those pictures would need to be converted into something a computer can understand, i.e. data.

The next two blocks in the diagram help with this very problem. Data preprocessing can be simply understood as analysing what the data represents, removing any duplicates or inconsistencies (a process called data cleaning), improving its clarity, and transforming it to be compatible with the model (data transformation). If training a model is equivalent to making food then data preprocessing is gathering all the ingredients, ensuring they haven’t gone bad and chopping them to an appropriate size. Just as a chef follows a recipe rather than tossing ingredients haphazardly into a pan, so too must we meticulously prepare our components.

This is where feature engineering comes into play. The idea is to take our transformed data and learn patterns from it, reduce its size to make training more efficient, and so on. The most important part of this process is Feature Representation. Let’s say you want to train a model to identify a heart in medical images. You will need to supply it with pictures of various organs and label them (e.g., ‘heart’, ‘lung’, ‘liver’). Then, you’ll need to convert each picture into numbers that a computer can understand.

This conversion process, where you decide how the image information will be numerically encoded, is Feature Representation. You could represent each picture by its individual pixel values – essentially, a grid of numbers where each number corresponds to the color/intensity of a tiny dot in the image. Alternatively, and often more effectively, you can focus on abstracting features like textures and edges. While modern AI models learn highly complex visual patterns automatically, these fundamental elements like edges often play a crucial role in defining the distinct shapes and boundaries of objects, which helps the model to recognize them effectively.

The final block represents the models that can be selected. There is no ‘one size fits all’ model; each has its own strengths and weaknesses. It is often dependent on the context that a certain model is chosen. This article will not delve into the intricacies of each model mentioned, nor the numerous others that exist, but it is important to understand the major differences between Machine Learning (ML) and Deep Learning (DL).

A crucial distinction between the two is captured perfectly with the following quote, ‘All DL models are ML models, but not all ML models are DL models.’ In other words, DL is a subset of ML that uses ‘neurons’ to simulate the working of our brains. DL is much more resource intensive in terms of computation power and data required to train the model satisfactorily. These models are also difficult to interpret owing to their ‘black box’ nature.

This nature of the DL Neural Networks is due to many neurons being connected to each other in layers. These layers, and the connections within them, involve millions or even billions of parameters (the weights and biases that define the strength of connections). This vast number of parameters and complex interconnections makes reasoning of a certain output very difficult to track. In contrast, ML models are simpler to understand as their logic for giving a certain output is based on more transparent algorithms whose workings are more easily inspectable by the engineer or data scientist.