Deep Learning Based Cardiac MRI Segmentation: Do We Need Experts?

Deep learning methods are the de facto solutions to a multitude of medical image analysis tasks. Cardiac MRI segmentation is one such application, which, like many others, requires a large number of annotated data so that a trained network can generalize well. Unfortunately, the process of having a large number of manually curated images by medical experts is both slow and utterly expensive. In this paper, we set out to explore whether expert knowledge is a strict requirement for the creation of annotated data sets on which machine learning can successfully be trained. To do so, we gauged the performance of three segmentation models, namely U-Net, Attention U-Net, and ENet, trained with different loss functions on expert and non-expert ground truth for cardiac cine–MRI segmentation. Evaluation was done with classic segmentation metrics (Dice index and Hausdorff distance) as well as clinical measurements, such as the ventricular ejection fractions and the myocardial mass. The results reveal that generalization performances of a segmentation neural network trained on non-expert ground truth data is, to all practical purposes, as good as that trained on expert ground truth data, particularly when the non-expert receives a decent level of training, highlighting an opportunity for the efficient and cost-effective creation of annotations for cardiac data sets.

Inisialisasi Otomatis Metode Level Set untuk Segmentasi Objek Overlapping pada Citra Panorama Gigi

<p class="Abstrak">Penelitian tentang segmentasi gigi individu telah banyak dilakukan dan memperoleh hasil yang baik. Namun, ketika dihadapkan kepada gigi overlap maka hal ini menjadi sebuah tantangan. Untuk memisahkan dua gigi overlap, maka perlu mengekstrak objek overlap terlebih dahulu. Metode level set banyak digunakan untuk melakukan segmentasi objek overlap, namun memiliki kelemahan yaitu perlu didefinisikan inisial awal metode level set secara manual oleh pengguna. Dalam penelitian ini diusulkan strategi inisialisasi otomatis pada metode level set untuk melakukan segmentasi gigi overlap menggunakan Hierarchical Cluster Analysis (HCA) pada citra panorama gigi. Tahapan strategi yang diusulkan terdiri dari preprocessing dimana di dalamnya ada proses perbaikan, rotasi dan cropping citra, dilanjutkan proses inisialisasi otomatis menggunakan algoritma HCA , dan yang terakhir segmentasi menggunakan metode level set. Hasil evaluasi menunjukkan bahwa strategi yang diusulkan berhasil melakukan inisialisasi secara otomatis dengan akurasi 73%. Hasil evaluasi segmentasi objek overlap cukup memuaskan dengan rasio misclassification error 0,93% dan relative foreground area error 24%. Dari hasil evaluasi menunjukkan bahwa strategi yang diusulkan dapat melakukan inisialisasi otomatis dengan baik. Inisialisasi yang tepat menghasilkan segmentasi yang baik pada metode level set.</p><p><em><strong><br /></strong></em></p><p><em><strong>Abstract</strong></em></p><p class="Judul2"><em>Individual teeth segmentation has done a lot of the recent research and obtained good results.</em><em> W</em><em>hen faced with overlapping teeth, this is quite challenging. To separate overlapping teeth, it is necessary to extract the overlapping object first. </em><em>The l</em><em>evel set method is widely used to segment overlap objects, but it has a limitation that needs to define the initial</em><em> </em><em>level set method manually by the user. This research proposes an automatic initialization strategy for the level set method to segment overlapping teeth using Hierarchical Cluster Analysis on dental panoramic radiograph images. The proposed strategy stage consists of preprocessing </em><em>where</em><em> there </em><em>are</em><em> several process</em><em>es</em><em> of enhancement, rotation</em><em>,</em><em> and cropping of the image, Then the automatic initialization process uses the HCA algorithm and the last is segmentation using the level set method. The evaluation results show that the proposed strategy is successful in carrying out automatic initialization with an accuracy of 73%. The results of the overlap object segmentation evaluation are satisfactory with a misclassification error ratio of 0.93% and a relative foreground area error of 24%. The evaluation results show that the proposed strategy can carry out automated initialization well. Proper initialization results can perform good segmentation of the level set method.</em></p><p><em><strong><br /></strong></em></p>

Automatic Segmentation of Left Ventricle in Echocardiography Based on YOLOv3 Model to Achieve Constraint and Positioning

Cardiovascular disease (CVD) is the most common type of disease and has a high fatality rate in humans. Early diagnosis is critical for the prognosis of CVD. Before using myocardial tissue strain, strain rate, and other indicators to evaluate and analyze cardiac function, accurate segmentation of the left ventricle (LV) endocardium is vital for ensuring the accuracy of subsequent diagnosis. For accurate segmentation of the LV endocardium, this paper proposes the extraction of the LV region features based on the YOLOv3 model to locate the positions of the apex and bottom of the LV, as well as that of the LV region; thereafter, the subimages of the LV can be obtained, and based on the Markov random field (MRF) model, preliminary identification and binarization of the myocardium of the LV subimages can be realized. Finally, under the constraints of the three aforementioned positions of the LV, precise segmentation and extraction of the LV endocardium can be achieved using nonlinear least-squares curve fitting and edge approximation. The experiments show that the proposed segmentation evaluation indices of the method, including computation speed (fps), Dice, mean absolute distance (MAD), and Hausdorff distance (HD), can reach 2.1–2.25 fps, 93.57 ± 1.97 % , 2.57 ± 0.89 mm, and 6.68 ± 1.78 mm, respectively. This indicates that the suggested method has better segmentation accuracy and robustness than existing techniques.

ID-Seg: An Accurate and Reliable Infant Deep learning Segmentation Framework for Limbic Structures

Early post-natal period brain magnetic resonance imaging (MRI) is becoming a common non-invasive approach to characterize the impact of prenatal exposures on neurodevelopment and to investigate early biomarkers for risk. Limbic structures are particular of interest in psychiatric disorder related research. Despite the promise of infant neuroimaging and the success of initial infant MRI studies, assessing limbic structure and function remains a significant challenge due to low inter-regional intensity contrast and high curvature (e.g. hippocampus). Of note, the agreement between existing automatic techniques and manual segmentation remains either untested or poor particularly for the amygdala and hippocampus. In this work, we developed an accurate (based on three segmentation evaluation metrics), reliable and efficient infant deep learning segmentation framework (ID−Seg) to address the aforementioned challenges. Specifically, we leveraged a large dataset of 473 infant MRI scans to train ID−Seg and then evaluated ID−Seg performance on internal (n=20) and external datasets (n=10) with manual segmentations. Compared with a state-of-the-art segmentation pipeline, we demonstrated that ID−Seg significantly improved the segmentation accuracy of limbic structures (hippocampus and amygdala) in newborn infants. Moreover, in a small, proof−of−concept analysis, we found that ID-Seg derived morphometric measures yield strong brain−behavior associations. As such, our ID-Seg may improve our capacity to efficiently measure MRI−based brain features relevant to neuropsychological development, and ultimately advance the success of quantitative analyses on large-scale datasets.

Semantic Segmentation of Aerial Imagery via Split-Attention Networks with Disentangled Nonlocal and Edge Supervision

In this work, we propose a new deep convolution neural network (DCNN) architecture for semantic segmentation of aerial imagery. Taking advantage of recent research, we use split-attention networks (ResNeSt) as the backbone for high-quality feature expression. Additionally, a disentangled nonlocal (DNL) block is integrated into our pipeline to express the inter-pixel long-distance dependence and highlight the edge pixels simultaneously. Moreover, the depth-wise separable convolution and atrous spatial pyramid pooling (ASPP) modules are combined to extract and fuse multiscale contextual features. Finally, an auxiliary edge detection task is designed to provide edge constraints for semantic segmentation. Evaluation of algorithms is conducted on two benchmarks provided by the International Society for Photogrammetry and Remote Sensing (ISPRS). Extensive experiments demonstrate the effectiveness of each module of our architecture. Precision evaluation based on the Potsdam benchmark shows that the proposed DCNN achieves competitive performance over the state-of-the-art methods.

Development of a lung segmentation algorithm for analog imaged chest X-Ray: preliminary results

Analog X-Ray radiography is still used in many underdeveloped regions around the world. To allow these populations to benefit from advances in automatic computer-aided detection (CAD) systems, X-Ray films must be digitized. Unfortunately, this procedure may introduce artefacts which may severely impair the performance of such systems. This work investigates the impact digitized images may cause to deep neural networks trained for lung (semantic) segmentation on digital x-ray samples. While three public datasets for lung segmentation evaluation exist for digital samples, none are available for digitized data. To this end, a U-Net architecture was trained on publicly available data, and used to predict lung segmentation on a newly annotated set of digitized images. Our results show that the model is capable to effectively identify lung segmentation at digital X-Rays with a high intra-dataset (PR AUC: 0.99) and cross-dataset (PR AUC: 0.99) performances on unseen test data. When challenged against analog imaged films, the performance is substantially degraded (PR AUC: 0.90). Our analysis further suggests that the use of maximum F1 and precision-recall AUC (PR AUC) measures are not informative to identify segmentation problems in images.