Instructor: Dr. Najeeb Khan

Instructor: Dr. Zahra Azizi

Instructors: Sébastien Quetin and Yujing Zou

Speaker: Dr. Julien Cohen-Adad

How does pre-processing of medical images affect the model generalizability, especially for training multi-modal medical imaging tasks?

Speaker: Dr. Marjory Jolicoeur

Instructor: Kiri Stern

Speaker: Dr. Maaike Berbee

To date prediction models are slowly finding their way into clinical practice. They are for example being used to facilitate shared decision making and, in the Netherlands, to selectively offer proton therapy to those patients who are expected to benefit from proton therapy. In this talk I will show how we validated a model to predict 2-year mortality in order to implement a model-based selection approach for proton therapy in esophageal cancer patients. Besides going through this example we will discuss the questions “Why do we need models of treatment effects in medicine” and “How to get models into wide use in the clinic”.

Speaker: Dr. Ane Appelt

Modern radiotherapy uses complex dose planning software to estimate and optimise the radiation dose delivered to the individual patient, based on 3D anatomical imaging. Consequently, we have detailed 3D data on the distribution of radiation dose to different parts of the body for every patient treated for radiotherapy – and, if we can predict who will get tumour control and/or toxicity, we can adjust this dose distribution to achieve the optimal ratio between the two. However, most traditional methods for outcome prediction only use very reduced data representations as input to the prediction models (e.g. ‘mean dose to an organ’), not the full 3D data. I’ll describe how we might apply CNN-based methods for classification to the problem of identifying ‘toxic’ radiation dose distributions; treating the dose information as just another type of image. But I’ll also discuss some of the pitfalls of this approach, and the large – and still unsolved – step of feeding the model output back into the radiotherapy treatment planning process.

Instructor: Hunter Buckhorn

Instructors: Sébastien Quetin and Yujing Zou

Speaker: Dr. Marianne Aznar

Speaker: Dr. Udunna Anazodo

Instructor: Dr. Mohammad Reza Mehebbian

Instructor: Dr. Mohammad Reza Mehebbian

Speakers: Dr. Karl Otto, Dr. Joshua Giambattista, Carter Kolbeck and Jon Giambattista (Limbus AI); Dr. Matthijs Kruis (Philips)

Speaker: Dr. Andreu Badal

What is the FDA approval process when deciding whether or not to commercialize a research product?

Instructor: Hunter Buckhorn

Instructors: Sébastien Quetin and Yujing Zou

Speaker: Dr. Anne Martel

Obtaining large datasets with detailed annotations for medical imaging AI projects is a time consuming and expensive process as it usually requires the input of expert radiologists and pathologists. Collecting data to train outcome prediction models is even more challenging as the number of patients with both imaging and follow up data may be small, and only weak labels are available. This talk will describe several semi-supervised and self-supervised approaches which can make more efficient use of small, weakly labelled datasets. The focus will be on digital pathology but the methods described are applicable any medical imaging modality.

Instructor: Dr. Mohammad Reza Mehebbian

Instructor: Dr. Katie Ovens

Speaker: Dr. Sarah Mattonen

In this presentation Dr. Mattonen will introduce the concept of radiomics and describe the overall workflow. She will outline the different types of radiomic features and walk through an example of how radiomic features are calculated. Best practices for conducting radiomic studies will also be discussed as well as some of the available open-source tools for performing radiomic studies. She will end by describing some of the work her lab is doing on applying radiomics to multi-modal imaging for outcome prediction in lung cancer.

Speaker: Dr. Heather M. Whitney

Characterization of case-based repeatability can complement AI performance metrics. We investigated case-based repeatability of classification of breast lesions as malignant or benign using multi-modality human-engineered radiomic features extracted from both dynamic contrast-enhanced magnetic resonance (DCE-MR) and full field digital mammography (FFDM) images using a database of 78 lesions. Twenty-eight DCE-MR features describing shape, morphology, texture, and kinetics of contrast enhancement and 32 FFDM features describing size, shape, margin, and texture were extracted after the lesions had been segmented using previously established methods. FFDM features for each lesion were averaged across all views available. Case-based repeatability was investigated for three scenarios: (1) DCE-MR features alone, (2) FFDM features alone, and (3) all features (i.e., multi-modality). The case-based repeatability profile for each scenario was developed by separating features into training and test folds by case using a 0.632 bootstrap with 200 iterations. A random forest classifier from each training fold was applied to each test fold, resulting in the posterior probability of malignancy for each case, called the case-based output (CBO). The CBO was scaled to 50% prevalence. The median CBO and the width of its 95% CI (95CICBO) were determined for each case across bootstrap folds. 95CICBO was also reviewed for each pairwise comparison of features used for classification, assessed using the coefficient of determination. Using multi-modality features somewhat modified the case-based repeatability profile for the lesions. It also increased the repeatability of some cases but decreased it for others. The 95CICBO of individual cases demonstrated little correlation between modalities. Classification of breast lesions using multi-modality human-engineered radiomic features may change case-based repeatability.

Information regarding MIDRC, the Medical Imaging and Data Resource Center, will also be presented.

Instructors: Dr. Farhad Maleki and Hossein Jafarzadeh

Instructor: Dr. Ibrahim Chamseddine