I am a medic, so excuse the banality of my question and my current lack of technical understanding on the matter. I am planning to conduct a systematic review on medical literature, in short: convolutional neural networks (CNNs) used for neuroradiology (CT imaging) and their relevance/prospects for clinical use.
A systematic review has to be systematic. Developing a scientific reproducible search protocol for finding relevant papers on CNNs applied to CT imaging has turned out tricky. "Convolution" defines my inclusion criteria: only ML models built on CNN architecture. How to semantically and scientifically define "CNN architecture" when the variety of models, (sometimes overlapping) terms and 'nicknames' for different solutions is vast and ever-expanding?
Carrying out the search for such a quest requires unambiguous definitions. If it were easy, I could go with: "convolut*/ConvNet/CNN" as search terms/keywords for "convolutional (NNs)". However, I am afraid is not a wide enough scope as convolution operations are used in various ML models (thus considered CNN-architecture based?), and spelling out "convolutional (neural network)" is not done on the paper as it is probably considered obvious? And here comes the problem: As I am not familiar with all existing backbone convolution networks and their names (ShuffleNet, Inception, GoogLeNet etc), it is impossible to systematically add them all to the search protocol and I might miss them, if I only search for "convolut*". See?
To start with high sensitivity (on the cost of specificity) I have built my search from top to bottom on the AI tree: I have included all following search terms from AI > ML > DL > NNs > CNNs and combined them with relevant medical terms. The search (several databases combined) yielded > 5.000 papers, and my plan was to filter (by hand) the papers that have have truly used a CNN. I excluded the obvious no's, and now have a few hundred papers left, but:
Inclusion and exclusion criteria based on algorithm details is hard:
First, my search is vast (terms from AI downwards to CNN) and some papers report the ML methods in unspecific terms on Title/Abstract: e.g. "We developed a new ML model to predict stroke". I cannot include this paper until consulting full text.
Second, in full text, convolution/CNN/ConvNet is not necessarily explicitly stated (considered obvious?): a quick Cntrl+F: "conv" doesn't catch a thing. However, they might describe the algorithm they've used by its name e.g. AlexNet, which is a CNN, or ResNet, some of which are often (?) at least partly convolution-based...
How to define "convolutional" neural networks... Can you help me? CNN is an ANN. The AI tree gets a bit messy at the level of CNNs: RNNs, ResNets, AEs, GANs, TransfLearn... Methods and terminology overlap and evolve constantly. Where do I draw the lines for CNNs?
In theory I should go through all these few hundred papers mentioning any ML method that might be convolutional NN-based and read them through carefully (I cannot afford consulting the code, I must rely on text descriptions). The problem is the number of papers which makes it almost impossible to go through in such detail to understand the model architecture. Automation for keyword hunting is also hard, as I don't have a set of all possible CNN keywords.
I am a medic and my abilities in understanding ML terminology is clearly limited. What do you think of my problem ? How should I approach it ?