Machine learning to guide cancer management

Oncology, health informatics, machine learning, computer vision, biostatistics and genomic statistics, mathematical modelling, radiology, radiation oncology, pathology, patient partners

Institutions

Campus Montréal : Université de Montréal, Polytechnique Montréal, HEC Montréal

Research centers and universities: Centre de recherche du CHUM (CRCHUM), Montreal Institute for Learning Algorithms (MILA), McGill, Concordia

University hospitals: Centre Hospitalier Université de Montréal (CHUM), Centre Hospitalier Universitaire Ste-Justine (CHUSJ)

Potential Partners

Cancer societies: Terry Fox Research Institute (TFRI), Digital Health Discovery Platform (DHDP), Cancer Research Society (CRS), Canadian Cancer Research Alliance (CCRA), Canadian Cancer Society Research Institute (CCSRI), OncoTech, Clinical Research in Oncology (Q-CROC), Exactis

Data science research institutes: Observatoire international sur les impacts sociétaux de l’IA et du numérique (OBVIA), Institute for Data Valorization (IVADO)

Funding agencies and research networks: CIFAR, Canada Research Chairs, Canadian Institutes of Health Research (CIHR), Fonds de la recherche du Québec en Santé (FRQS), Réseau Bio-imagerie Québec (QBIN), Génome Québec, Génome Canada

Patient partnership and hospital foundation: Centre of Excellence on Partnership with Patients and the Public, Fondation du CHUM

Research consortiums, institutes, technology accelerators and incubators:

Consortium Québécois sur la Découverte du Médicament (CQDM), IRICoR,

TransMedTech, Consortium Québécois pour recherche industrielle et innovation en technologies médicales (MEDTEQ), MITACS-Cluster Accelerate

List of hospitals linked by the CODA-19 CIHR-funded initiative:

1. CHUM
2. Jewish General Hospital
3. McGill University Health Centre
4. CIUSSS de l’Estrie
5. Hôpital du Sacré-Coeur
6. CHU de Québec – Université Laval
7. CHU Sainte-Justine
8. CISSS de Chaudière-Appalaches
9. Ottawa Hospital

Les noms suivants ont été proposés par la communauté et les personnes mentionnées ci-dessous ont accepté d’afficher publiquement leur nom. Notez cependant que tous les noms des professeur.e.s (qu’ils soient affichés publiquement ou non sur notre site web) seront transmis au comité conseil pour l’étape d’identification et de sélection des thèmes stratégiques. Notez également que les personnes identifiées durant l’étape de consultation n’ont pas la garantie de recevoir une partie du financement. Cette étape sert avant tout à présenter un panorama du domaine, incluant les personnes pertinentes et non à monter des équipes pour les programmes-cadres.

• Anne-Marie Mes-Masson
• Aurélie Labbé
• An Tang
• Gerald Batist
• Houda Bahig
• Michael Chassé
• Morgan Craig
• Sahir Bhatnagar
• Samuel Kadoury
• Simon Turcotte
• Guy Wolf

To use machine learning to guide cancer management, there is a need to: 1) build a network linking hospitals caring for cancer patients, 2) distribute data annotations, 3) investigate cancer analytics addressing unmet needs in clinical care. The following research frameworks respectively mirror these needs:

Aggregating data of numerous patients from several institutions will be required to include a diverse population. However, because of concerns regarding data privacy and the challenges of pooling data in central repositories within or between Canadian provinces, it is anticipated that federated learning will be the preferred approach to train models by weights transfer. This will provide an opportunity to investigate and design frameworks that work with different institutes, equipment, and patient distributions. Different frameworks will provide trade-offs between fairness, privacy, safety, and robustness. We propose to build this strategic application on the CIHR- and FRQS-funded CODA-19 network developed during the COVID-19 pandemic to link in near real-time nine hospitals. The conceptual and theoretical insights that we will acquire through this work will be applicable to future work on federated learning at the national level with the Digital Health Discovery Platform (DHDP).

Training models using supervised or semi-supervised learning requires labelled data. This strategic proposal to build a multi-institutional and multi-omics oncological dataset will require labelling by experts in different hospitals and research centers to provide valid ground truths (i.e. reference standards). These annotations may include bounding boxes, labels, and segmentation masks on radiological images or digitized histopathology slides; classes such as tumor subtypes; and dates such as time to tumor recurrence or survival time. Whether we consider adopting existing software solutions or developing our own, we must ensure that our labels can be shared with the data science community and compatible with future national research networks (such as DHDP or CIHR-funded Network of Networks).

Data analytics may be diagnostic, descriptive, predictive, or prescriptive. Multiple lesions may coexist in the same organ. Also, tumors are heterogeneous and respond differently to treatment over time, even in a given patient. At a high level, we need to pursue:

• Diagnostic cancer analytics to classify individual lesions (in the past)
• Descriptive cancer analytics to assess treatment response (in the present
• Predictive cancer analytics to anticipate what will happen (in the future)
• Prescriptive cancer analytics to recommend preferred treatment strategies (in the present) to affect outcomes (in the future)

(pas de documentation complémentaire pour le moment)