Medical image analysis with deep learning for computer-aided diagnosis in screening ; Analyse d'images médicales par apprentissage profond pour le diagnostic assisté par ordinateur dans un contexte de dépistage
You can manage bookmarks using lists, please log in to your user account for this.
Media type:
E-Book
Title:
Medical image analysis with deep learning for computer-aided diagnosis in screening ; Analyse d'images médicales par apprentissage profond pour le diagnostic assisté par ordinateur dans un contexte de dépistage
Published:
[Erscheinungsort nicht ermittelbar]: HAL CCSD, 2021
Language:
English
Origination:
University thesis:
Dissertation, HAL CCSD, 2021
Footnote:
Description:
Computer-aided medical image analysis is essential to support clinicians in diagnosis, prognosis and therapy-related decisions through fast, repeatable and objective measurements made by computational resources. In particular, the latest development of artificial intelligence applied to diagnosis and screening represents a promising perspective. In this thesis, we addressed the current limitations of traditional computer-aided diagnosis (CAD) systems by providing efficient and fully-automated deep learning methods towards better interaction-free and more personalized medical care. In the contexts of breast cancer and diabetic retinopathy screening, we investigated three main challenges associated with computer-assisted medical image analysis: (1) identification and segmentation of lesions from high-resolution images, (2) multi-view information fusion for improved diagnosis, and (3) longitudinal prediction of severity grade changes. Our initial contribution to the first challenge was to propose an end-to-end mass segmentation pipeline that exploits long-range multi-scale spatial context through a cascade of convolutional encoder-decoders embedding the auto-context paradigm. Then, as a second contribution, we proposed a two-stage framework combining a deep coarse-scale mass localization involving a multi-scale fusion strategy and a fine-scale mass segmentation. The second challenge was addressed by fusing information arising from two standard mammography views, namely craniocaudal (CC) and mediolateral-oblique (MLO). Two methods were proposed towards this goal. First, a novel approach based on multi-task learning was introduced, combining mass classification with dual-view mass matching between CC/MLO mammograms. Then, we applied a label-efficient deep active learning approach that exploits dual-view consistency to mitigate the labeling workload of clinicians. These methods demonstrate the effectiveness of integrating multi-view information for detection or segmentation purposes. For the last challenge, we ...