L’équipe ARMEDIA organise un séminaire sur les techniques d’apprentissage profond appliquées à l’imagerie médicale, ouvert à tout le labo,
le jeudi 4 mai, à 14h en H218 (bâtiment Etoile, 2è étage) avec lien visioconférence avec le site de NanoInnov à Saclay.
Titre : « Detection and quantification of infiltrative lung diseases using deep learning techniques »
Intervenant : M. Sebastian Tarando, doctorant TSP
Biographie : Sebastian TARANDO a obtenu son diplôme d’Ingénieur en Electronique de l’Université de Buenos Aires (UBA), Argentine en 2015. Il effectue actuellement sa thèse de doctorat au département ARTEMIS de TSP sous la direction de Catalin FETITA, sur une thématique liée à la détection de pathologies en imagerie pulmonaire. Ses intérêts de recherche portent sur les approches par apprentissage profond et les réseaux neuronaux convolutionnels.
Résumé :
The infiltrative lung diseases (ILD) are a class of irreversible, non-neoplastic lung pathologies requiring regular follow-up with CT-scan imaging. Quantifying the evolution of the patient status imposes the development of automated classification tools for the lung texture. For the large majority of computer-aided diagnosis (CAD) systems, such classification relies on a two-dimensional analysis of axial CT images.
The inherent issue of ILD classification task is the great overlap between the texture pattern appearance which often makes the task an ontological problem leading to errors in the classification, i.e., mixing different types of lung patterns or overestimating the pathological areas. Recently, the use of deep leaning techniques, especially Convolutional Neural Networks (CNNs), has shown great improvements with respect to traditional methods (based on handcrafted features), emerging as the new state of the art for visual tasks and the reference for further comparisons with new approaches. Deep learning methods have the advantage (over traditional CAD approaches) that the features on which relies the classification are learned directly from the image dataset. This leads apparently to a better image pattern discrimination than when using handcrafted features. Note that CNNs particularly have proven to succeed in natural image classification tasks.
In this study we investigate the benefit of using deep learning approaches for ILD classification in conjunction with or versus traditional methods.
In a previous research, a fully-3D ILD classification approach was developed exploiting a multi-scale morphological analysis which showed good performance in detecting diseased areas, but with a major drawback consisting of sometimes overestimating the pathological regions and mixing different types of lung patterns.
As a first investigation, we propose here a combination of this CAD system with the classification outcome provided by a convolutional network, specifically tuned-up in order to increase the specificity of the classification and the confidence to diagnosis, but trained on a reduced dataset. Here, the objective of using a deep learning approach is to achieve a better regularization of the classification output (because of a deeper insight into a given pathological class over a large series of samples) where the previous system is extra-sensitive due to the multi-scale response on patient-specific, localized patterns). A preliminary evaluation of the combined approach on a 10-patient database with various lung pathologies, showed a sharp increase of true detections.
In a second investigation having the objective of improving the performance and reducing the computational cost, we explore the CNN classification approach alone. We propose a cascade of an existing state-of-the-art CNN-based CAD system (T-CNN) achieving a first classification of less correlated patterns, with a second classification provided by a texture matching method (also CNN based) for those classes that are highly correlated (namely Fibrosis and Ground Glass). The combined approach was tested on a 13 patient database of various lung pathologies, showing an increase of 10% in True Positive Rate (TPR) with respect to the best suited state-of-the-art CNN for this task.
The talk is concluded with a critical analysis of the approaches presented and an outline of future research directions, followed by an open discussion with the audience.