RESUMEN
Weakly-supervised learning (WSL) has been proposed to alleviate the conflict between data annotation cost and model performance through employing sparsely-grained (i.e., point-, box-, scribble-wise) supervision and has shown promising performance, particularly in the image segmentation field. However, it is still a very challenging task due to the limited supervision, especially when only a small number of labeled samples are available. Additionally, almost all existing WSL segmentation methods are designed for star-convex structures which are very different from curvilinear structures such as vessels and nerves. In this paper, we propose a novel sparsely annotated segmentation framework for curvilinear structures, named YoloCurvSeg. A very essential component of YoloCurvSeg is image synthesis. Specifically, a background generator delivers image backgrounds that closely match the real distributions through inpainting dilated skeletons. The extracted backgrounds are then combined with randomly emulated curves generated by a Space Colonization Algorithm-based foreground generator and through a multilayer patch-wise contrastive learning synthesizer. In this way, a synthetic dataset with both images and curve segmentation labels is obtained, at the cost of only one or a few noisy skeleton annotations. Finally, a segmenter is trained with the generated dataset and possibly an unlabeled dataset. The proposed YoloCurvSeg is evaluated on four publicly available datasets (OCTA500, CORN, DRIVE and CHASEDB1) and the results show that YoloCurvSeg outperforms state-of-the-art WSL segmentation methods by large margins. With only one noisy skeleton annotation (respectively 0.14%, 0.03%, 1.40%, and 0.65% of the full annotation), YoloCurvSeg achieves more than 97% of the fully-supervised performance on each dataset. Code and datasets will be released at https://github.com/llmir/YoloCurvSeg.
RESUMEN
PURPOSE: Accurately segmenting curvilinear structures, for example, retinal blood vessels or nerve fibers, in the medical image is essential to the clinical diagnosis of many diseases. Recently, deep learning has become a popular technology to deal with the image segmentation task, and it has obtained remarkable achievement. However, the existing methods still have many problems when segmenting the curvilinear structures in medical images, such as losing the details of curvilinear structures, producing many false-positive segmentation results. To mitigate these problems, we propose a novel end-to-end curvilinear structure segmentation network called Curv-Net. METHODS: Curv-Net is an effective encoder-decoder architecture constructed based on selective kernel (SK) and multibidirectional convolutional LSTM (multi-Bi-ConvLSTM). To be specific, we first employ the SK module in the convolutional layer to adaptively extract the multi-scale features of the input image, and then we design a multi-Bi-ConvLSTM as the skip concatenation to fuse the information learned in the same stage and propagate the feature information from the deep stages to the shallow stages, which can enable the feature captured by Curv-Net to contain more detail information and high-level semantic information simultaneously to improve the segmentation performance. RESULTS: The effectiveness and reliability of our proposed Curv-Net are verified on three public datasets: two color fundus datasets (DRIVE and CHASE_DB1) and one corneal nerve fiber dataset (CCM-2). We calculate the accuracy (ACC), sensitivity (SE), specificity (SP), Dice similarity coefficient (Dice), and area under the receiver (AUC) for the DRIVE and CHASE_DB1 datasets. The ACC, SE, SP, Dice, and AUC of the DRIVE dataset are 0.9629, 0.8175, 0.9858, 0.8352, and 0.9810, respectively. For the CHASE_DB1 dataset, the values are 0.9810, 0.8564, 0.9899, 0.8143, and 0.9832, respectively. To validate the corneal nerve fiber segmentation performance of the proposed Curv-Net, we test it on the CCM-2 dataset and calculate Dice, SE, and false discovery rate (FDR) metrics. The Dice, SE, and FDR achieved by Curv-Net are 0.8114 ± 0.0062, 0.8903 ± 0.0113, and 0.2547 ± 0.0104, respectively. CONCLUSIONS: Curv-Net is evaluated on three public datasets. Extensive experimental results demonstrate that Curv-Net outperforms the other superior curvilinear structure segmentation methods.