AI workflow: the 4 components

We will use an example case of a new AI tool, Mesonet developed by Owkin, whose results were recently published in Nature. This tool was built to better understand malignant mesothelioma – an aggressive cancer type – and improve predicting prognosis of the disease. Currently, this type of cancer is diagnosed via extensive manual labor of a pathologist by scoring histological criteria. Getting a better impression of its prognosis and treatment strategy is an extremely difficult task that could benefit from using artificial intelligence.

AI workflow

The steps of building AI applications for predicting the development of diseases like malignant mesothelioma are illustrated in figure 1.

Figure 1. AI Workflow

Step 1. Digitalise and collect data

Collecting and storing data is one of the most important steps of the AI workflow. In many areas of healthcare, the data first needs to be digitised and made interoperable before it can be used for learning. The selected data needs to be formatted, cleaned and sampled. This step in the AI workflow is also key for identification of dataset fairness, which is the process of understanding bias introduced by the data and ensures a model with equitable predictions across all demographic groups.

Example Mesonet: Digitize and collect images of tumours. Set up access to validation group database and Cancer Genome Atlas Remove privacy-sensitivity, repair errors and sample data.

Step 2. Transform and build model

Data can exist in a wide variety of sizes and formats and needs to be harmonised (scaled, decomposed and aggregated) into a single format that is easily processed by an AI system. This requires an iterative process of 1) preparing the data for conversion, 2) converting the data to the target format and 3) evaluating the formatted data to identify unusable records.

Example Mesonet: 1) Categorize pixels as foreground (matter) or background with U-Net neural network 2) Divide matter into smaller images (“tiles”) 3) Obtain over 2,048 relevant features from each tile using a deep neural network

Step 3. Build and train

Having organised and well formatted datasets, neural network models then can be trained and evaluated with a training and test dataset, respectively. This element of training is key for ‘intelligence’ in AI, as learning makes the system smart. Training involves a repetitive set of steps executing mathematical functions on the data and is designed to identify a desired response/result with a high level of probability. The results are then evaluated for accuracy and fairness. The mathematical functions are modified and repeated by applying the updates to the same data set until a high level of accuracy and fairness is reached.

Example Mesonet: Trained and test model on 2,981 slides from 2,981 mesothelioma patients 2,300 training slides and 681 evaluation slides. Test robustness on independent dataset.

Step 4. Execute

Upon execution, the trained and refined AI models are deployed and used to make decisions. During this phase the accuracy of the model is also evaluated continually. The execution process is reanalyzed to make sure that the system is meeting expectations and to provide a feedback loop for improvement.

Example Mesonet: Generate a score for each tile, weighing the sum between all 2,048 features of the tile using a model consisting of a convolutional one-dimensional layer. Calculate a prediction from the scores of the tiles using multi-layer perception classifier.

Step 5. Translate the action

Use the insights provided to help decision making and improve our understanding of the disease.

Example Mesonet: The outcomes of the application were compared with standard scoring methods. The tool aids to identify new features of the tumor, is able to link these to overall prognosis and thereby offer a more accurate prediction for patient survival.

Bringing AI research into practice