Retinal Blood Vessel Segmentation Using a Fully Convolutional Network – Transfer Learning from Patch- to Image-Level

Abstract

Fully convolutional networks (FCNs) are well known to provide state-of-the-art results in various medical image segmentation tasks. However, these models usually need a tremendous number of training samples to achieve good performances. Unfortunately, this requirement is often difficult to satisfy in the medical imaging field, due to the scarcity of labeled images. As a consequence, the common tricks for FCNs' training go from data augmentation and transfer learning to patch-based segmentation. In the latter, the segmentation of an image involves patch extraction, patch segmentation, then patch aggregation. This paper presents a framework that takes advantage of all these tricks by starting with a patch-level segmentation which is then extended to the image level by transfer learning. The proposed framework follows two main steps. Given a image database D, a first network NP is designed and trained using patches extracted from D. Then, NP is used to pre-train a FCN NI to be trained on the full sized images of D. Experimental results are presented on the task of retinal blood vessel segmentation using the well known publicly available DRIVE database.

CNN‐based fully automatic wrist cartilage volume quantification in MR images: A comparative analysis between different CNN architectures

“…The tested U‐Net‐based CNNs were not only more efficient regarding the segmentation metrics but also regarding the training time. Accordingly, it has been reported that patch‐based approaches were less computationally effective than fully convolution networks with an image‐based learning approach 22,52 as they classify each pixel in an image separately. It is important to mention that, although the training time for the U‐Net optimized in nnU‐Net framework was longer than for the common U‐Nets, the overall time the authors spent for manual adaptation of the networks and for the grid search was much longer.…”

“…The noise was added as a Tensorflow layer directly to the CNN input. In addition to the training hyperparameters and as previously described, 22,42 number of layers (i.e., the depth of the CNN), order and presence of noise, dropout and batch normalization layers have been considered as adjustable architecture hyperparameters. Two variants of the U‐Nets were used: one with an original depth, that is, number of max‐poolings equal to 4, 30 and one with a reduced depth (Truncated U‐Net) that is, with three max‐poolings.…”

“…The most promising methods for fully automatic segmentation of complex biomedical structures are based on convolutional neural networks (CNNs) 20–22 . Several network architectures have shown excellent results for cartilage segmentation from MR images of knee 23 with Dice similarity coefficients (DSCs) up to 0.92, 24,25 while a performance for wrist cartilage segmentation has been poorly reported.…”

Using Deep Learning to Bootstrap Abstractions for Hierarchical Robot Planning

“…To handle such large environments without making any changes to the DNN, we use the standard approach of sliding windows with stride equal to window-width [3,13,27,34]. This crops the larger environment into pieces of the size of the training environments.…”

D-GaussianNet: Adaptive Distorted Gaussian Matched Filter with Convolutional Neural Network for Retinal Vessel Segmentation

“…Since in the area of biomedical image processing, it is common to work with a limited number of examples, different training strategies are required. Frequently, this issue is overcome by applying transferred learning [ 18 , 25 , 29 ] or using patches rather than full images for training [ 9 , 11 , 19 , 23 ], while a few papers have proposed hybrid techniques for feature extraction [ 12 , 34 ]. In this paper, a new method that incorporates distorted Gaussian Matched Filters with adaptive parameters as part of a Deep Convolutional Architecture is proposed.…”