Internship Opportunities at Imagia
Come train with us! At Imagia, we value teaching and learning from our peers. Our robust traineeship program is tailor-made to ensure each trainee gains invaluable on-site work experience and scientific coaching from world-class talent.
Radiomics discoveries are limited by the necessity of expert data curation from healthcare professionals (segmentation, insights on the outcome), and siloed research performed on local patient data at single hospitals preventing the emergence of sufficient clinical evidence.
Our goal is to leverage advances in image processing using semi-supervised and unsupervised deep learning methods, to automate enough of the curation of patient data for unsupervised radiomics discovery. Imagia has identified the following research avenues to help solve these challenges.
Neural architecture search: NAS is a family of algorithms that automatically discover network architecture best suited for a given dataset. Although the community has come a long way in recent years to scale these algorithms, they still can’t generalize to the drastically different sources of data that can be found in medical imaging: image size and resolution, number of examples, task complexity, etc.
Imagia already possesses its in house NAS toolkit, and is interested in extending its capacities. The intern must be comfortable with tensorflow.keras.
Exploring 3D architectures and transfer learning from 2D: if architectures for 2D RGB images (such as ImageNet) have been fine-tuned enough and are now mature enough to be used in the industry, their efficiency does not transfer easily in the 3D context such as CT scans. In our recent internal work, we were surprised by the behavior of classical models converted to 3D. We believe that applied research on 3D architecture will be key in improving deep learning efficiency over medical volumes.
The work realized through this internship is expected to be deployed through our Developer Platform and tested in a research environment from within the hospitals. The intern must be comfortable with tensorflow.keras.
Domain Adaptation and Transfer Learning: Machine learning models perform best when tested on target (test) domains that are similar to the source (train) domain that they are trained on. However, model generalization can be hindered when there is a significant covariate shift between the target and source domains. Medical datasets vary significantly due to inherent variability between modalities, patients, image acquisition protocols and medical centers. Domain adaptation and domain generalization paradigms address such a covariate shift by learning both domain agnostic and domain specific features. The main purpose of the project is to apply the various existing domain adaptation methods to Imagia’s clinical data, establish a guideline on when and how to use these methods and evaluate how they compare with other transfer-learning methodologies such as pre-training on larger datasets.
Points Clouds for CT scans: point cloud has been recently highlighted by a few scientific publications. They present very interesting properties such as a highly compressed representation for 3D and permutation invariance. To the best of our knowledge, there is no work using point cloud representation of CT scans and deep learning for medical tasks such as segmentation or classification.
Active Learning Fully automatic deep learning has become the state-of-the-art technique for many tasks including image acquisition, analysis and interpretation, and for the extraction of clinically useful information for computer-aided detection, diagnosis, treatment planning, intervention and therapy. However, the unique challenges posed by medical image analysis suggest that retaining a human end-user in any deep learning enabled system will be beneficial. We want to investigate the role that humans might play in the development and deployment of deep learning enabled diagnostic applications and focus on techniques that will retain a significant input from a human end-user.
Imagia already conducted some work on the domain. Currently, an appropriate framework exists and is ready to receive new models. The intern must be comfortable with tensorflow 2,0 and tensorflow.keras.
Self-supervised transfer learning. Supervised learning of deep neural networks has shown to be a powerful technique in training nonlinear function estimators with the caveat of requiring a lot of labeled data. In medical imaging however, labeled datasets are hard to come by and most available data are unlabeled. Self-supervised learning is a relatively recent technology which trains the model in a supervised fashion but on labels automatically generated using the data itself. While alleviating the need for manual annotations, the model can still learn meaningful representations of the data.
Generative 3D: generative models have been a breakthrough in machine learning, as they made a major step toward unsupervised learning. However, training generative models, even on 2D images in an acceptable resolution, remains a challenging task, especially for medical data that present high dimension (volumes), high resolution and high variability. Imagia has already conducted extensive work on the area. The intern will be in charge of extending the project by training using a dataset of 15K patients (about 45K 3D volumes) and to participate in the associated publication.