Use Cases in AI - Reasons to do AI with friends

Learning Objectives:
1. Considerations for dataset curation
2. Image segmentation (L3 selection and muscle mass)
3. Training and testing
4. Validation

Use Case Description
When assessing sarcopenia, the L3 level has to be detected manually on the whole body CT, the muscle bulk has to be segmented and corrections of these segmentations to remove adipose tissue have to be made manually. We will write an algorithm to automate detection of the L3 slice on CT and extract muscle mass for sarcopenia assessment.

Learning Objectives:
1. To Differentiate Renal Cell Carcinoma from Oncocytoma
2. Considerations for database development
3. Data annotation, Training, and testing
4. Validation
5. Final Output

Use Case Description
We propose to write an algorithm that will automatically detect and segment renal masses on CT. The segmented mass will then be further characterised using a radiomic approach, to differentiate benign from malignant masses.

Learning Objectives:
1. Despite best clinical imaging, prenatal evaluation cannot correctly predict good vs poor neurodevelopmental outcome in individual cases of isolated CCA
2. AI can and should be used in fetal MRI
3. Using AI we can study patterns of fetal brain connectivity and development beyond tractography, evaluating development and deviations possibly associated with disease
4. Applying AI to fetal MRI requires specific troubleshooting

Use Case Description Corpus callosum agenesis (CCA) is one of the most common brain malformations. Fetal MRI improves identification of parenchymal anomalies in CCA in relation to US. When isolated, CCA has a good prognosis in 70-80% of children. However, using currently available methods, we cannot correctly predict the fetuses at risk of neurodevelopmental delay. AI is a tool that can be used in fetal MRI. Peculiarities must however be taken into consideration, pertaining to the continuous changes of the developing brain, fetal and maternal movement, to name a few. Using DTI data, models of development of the fetal connectome can be developed to quantify deviations associated with disease and establish a vocabulary we can link to outcome. For this, all available data should be used. For validation of these models, it is essential to have a close relation to clinicians and a good follow up of patients.

Learning Objectives:
1. To identify malignant lesions on whole body MRI

Use Case Description
We propose to write an algorithm that will automatically detect malignant lesions on whole body MRI. We plan to use the large prospectively collected data from both STREAMLINE studies (STC and STL) together with the consensus reference data on confirmed sites of disease to train an ML algorithm to detect the primary and metastatic lesions.

Learning Objectives:
1. Detection, classification & segmentation of lesions in breast tissue
2. Data requirements in Machine Learning
3. Data annotation & validation in Machine Learning

Use Case Description
We propose to write an algorithm that will automatically detect malignant breast lesions. In this webinar we will explain to you, why diversity is crucial in the data set and what needs to be considered when annotating and validating data sets in the lab and clinically. Our machine learning algorithm will be able to detect, classify, and segment lesions in multi-modal and multi-parametric imaging data.

Learning Objectives:
1. To understand the difference between supervised and unsupervised learning
2. To appreciate the need for large-scale data sets from different sites and scanners
3. To understand how unsupervised algorithms can be validated

Use Case Description
Humans cannot identify complex patterns in CT data reliably. In the detection of fibrosis, relevant patterns need to be identified and we suspect that we don’t know all of them. We will write an algorithm that detects the relevant type of fibrosis and explain the difference between supervised and unsupervised learning.