Kibrom Berihu Girum1,2, Gilles Créhange3,4, Raabid Hussain3, Alain Lalande3,5. 1. ImViA Laboratory, University of Burgundy, Dijon, France. kibrom-berihu_girum@etu.u-bourgogne.fr. 2. Radiation Oncology Department, CGFL, Batiment I3M, 64b rue sully, 21000, Dijon, France. kibrom-berihu_girum@etu.u-bourgogne.fr. 3. ImViA Laboratory, University of Burgundy, Dijon, France. 4. Radiation Oncology Department, CGFL, Batiment I3M, 64b rue sully, 21000, Dijon, France. 5. Medical Imaging Department, CHU Dijon, Dijon, France.
Abstract
PURPOSE: To achieve accurate image segmentation, which is the first critical step in medical image analysis and interventions, using deep neural networks seems a promising approach provided sufficiently large and diverse annotated data from experts. However, annotated datasets are often limited because it is prone to variations in acquisition parameters and require high-level expert's knowledge, and manually labeling targets by tracing their contour is often laborious. Developing fast, interactive, and weakly supervised deep learning methods is thus highly desirable. METHODS: We propose a new efficient deep learning method to accurately segment targets from images while generating an annotated dataset for deep learning methods. It involves a generative neural network-based prior-knowledge prediction from pseudo-contour landmarks. The predicted prior knowledge (i.e., contour proposal) is then refined using a convolutional neural network that leverages the information from the predicted prior knowledge and the raw input image. Our method was evaluated on a clinical database of 145 intraoperative ultrasound and 78 postoperative CT images of image-guided prostate brachytherapy. It was also evaluated on a cardiac multi-structure segmentation from 450 2D echocardiographic images. RESULTS: Experimental results show that our model can segment the prostate clinical target volume in 0.499 s (i.e., 7.79 milliseconds per image) with an average Dice coefficient of 96.9 ± 0.9% and 95.4 ± 0.9%, 3D Hausdorff distance of 4.25 ± 4.58 and 5.17 ± 1.41 mm, and volumetric overlap ratio of 93.9 ± 1.80% and 91.3 ± 1.70 from TRUS and CT images, respectively. It also yielded an average Dice coefficient of 96.3 ± 1.3% on echocardiographic images. CONCLUSIONS: We proposed and evaluated a fast, interactive deep learning method for accurate medical image segmentation. Moreover, our approach has the potential to solve the bottleneck of deep learning methods in adapting to inter-clinical variations and speed up the annotation processes.
PURPOSE: To achieve accurate image segmentation, which is the first critical step in medical image analysis and interventions, using deep neural networks seems a promising approach provided sufficiently large and diverse annotated data from experts. However, annotated datasets are often limited because it is prone to variations in acquisition parameters and require high-level expert's knowledge, and manually labeling targets by tracing their contour is often laborious. Developing fast, interactive, and weakly supervised deep learning methods is thus highly desirable. METHODS: We propose a new efficient deep learning method to accurately segment targets from images while generating an annotated dataset for deep learning methods. It involves a generative neural network-based prior-knowledge prediction from pseudo-contour landmarks. The predicted prior knowledge (i.e., contour proposal) is then refined using a convolutional neural network that leverages the information from the predicted prior knowledge and the raw input image. Our method was evaluated on a clinical database of 145 intraoperative ultrasound and 78 postoperative CT images of image-guided prostate brachytherapy. It was also evaluated on a cardiac multi-structure segmentation from 450 2D echocardiographic images. RESULTS: Experimental results show that our model can segment the prostate clinical target volume in 0.499 s (i.e., 7.79 milliseconds per image) with an average Dice coefficient of 96.9 ± 0.9% and 95.4 ± 0.9%, 3D Hausdorff distance of 4.25 ± 4.58 and 5.17 ± 1.41 mm, and volumetric overlap ratio of 93.9 ± 1.80% and 91.3 ± 1.70 from TRUS and CT images, respectively. It also yielded an average Dice coefficient of 96.3 ± 1.3% on echocardiographic images. CONCLUSIONS: We proposed and evaluated a fast, interactive deep learning method for accurate medical image segmentation. Moreover, our approach has the potential to solve the bottleneck of deep learning methods in adapting to inter-clinical variations and speed up the annotation processes.