During the last century, continuous advances in biomedical imaging technologies gave rise to a wide variety of visual representations of the interior of living organisms at the organ, tissue, cellular and molecular level. Modalities such as x-ray, nuclear and molecular imaging, ultrasound, MRI and CT play a crucial role in clinical practice and basic life sciences research. Nowadays it is possible to process and understand this data thanks to the development of computational methods for the analysis of biological and medical images.
The aim of this research line is to develop novel machine learning methods to tackle some of the main problems in the field of biological and medical image analysis, including image registration, segmentation and classification, with particular focus on deep learning models. Our goal is to address fundamental machine learning problems such as domain adaptation, algorithmic fairness, learning with limited data and non-regular structures which arise in the context of image analysis for biology and medicine.
Post-DAE: a post-processing method to improve the anatomical plausibility of segmentation masks using denoising autoencoders.
Orthogonal Ensemble Networks for Image Segmentation: ining strategy for deep ensembles of segmentation networks that boosts discrimination and calibration performance.
Head CT Tools: a toolkit that includes a bunch of algorithms for preprocessing CT images & PyTorch based models for brain segmentation or skull reconstruction on craniectomy images.
ChronoRoot: a system for high-throughput plant root phenotyping composed of 3D printed hardware and deep learning methods for image analysis.
HybridGNet: a new hybrid graph convolutional neural network architecture for landmark-based anatomical segmentation.
AC-RegNet: a new method to regularize CNN-based deformable image registration by considering global anatomical priors in the form of segmentation masks.
BLAST: the Brain Lesion Analysis & Segmentation Tool, implements pre-processing pipelines and several CNN architectures (DeepMedic, UNet3D, etc) for lesion and anatomy segmentation in brain images using Lasagne (Theano).
GenderBiasX-Ray: a fairness study where we show that computer assisted diagnosis systems acquire gender bias when training datasets are highly imbalanced.