Automatic image segmentation based on synthetic tissue model for delineating organs at risk in spinal metastasis treatment planning

Accurate object localization facilitates automatic esophagus segmentation in deep learning

BACKGROUND

Currently, automatic esophagus segmentation remains a challenging task due to its small size, low contrast, and large shape variation. We aimed to improve the performance of esophagus segmentation in deep learning by applying a strategy that involves locating the object first and then performing the segmentation task.

METHODS

A total of 100 cases with thoracic computed tomography scans from two publicly available datasets were used in this study. A modified CenterNet, an object location network, was employed to locate the center of the esophagus for each slice. Subsequently, the 3D U-net and 2D U-net_coarse models were trained to segment the esophagus based on the predicted object center. A 2D U-net_fine model was trained based on the updated object center according to the 3D U-net model. The dice similarity coefficient and the 95% Hausdorff distance were used as quantitative evaluation indexes for the delineation performance. The characteristics of the automatically delineated esophageal contours by the 2D U-net and 3D U-net models were summarized. Additionally, the impact of the accuracy of object localization on the delineation performance was analyzed. Finally, the delineation performance in different segments of the esophagus was also summarized.

RESULTS

The mean dice coefficient of the 3D U-net, 2D U-net_coarse, and 2D U-net_fine models were 0.77, 0.81, and 0.82, respectively. The 95% Hausdorff distance for the above models was 6.55, 3.57, and 3.76, respectively. Compared with the 2D U-net, the 3D U-net has a lower incidence of delineating wrong objects and a higher incidence of missing objects. After using the fine object center, the average dice coefficient was improved by 5.5% in the cases with a dice coefficient less than 0.75, while that value was only 0.3% in the cases with a dice coefficient greater than 0.75. The dice coefficients were lower for the esophagus between the orifice of the inferior and the pulmonary bifurcation compared with the other regions.

CONCLUSION

The 3D U-net model tended to delineate fewer incorrect objects but also miss more objects. Two-stage strategy with accurate object location could enhance the robustness of the segmentation model and significantly improve the esophageal delineation performance, especially for cases with poor delineation results.

Feasibility of Augmented Reality for Pediatric Giant Supratentorial Tumors: A Report of Three Cases

Giant supratentorial brain tumors (GSBTs) in children are uncommon and extremely challenging entities unique to pediatric neurosurgery. Factors such as young patient age, need for urgent intervention, intraoperative blood loss, and ongoing raised intracranial pressure symptoms are examples of difficulties faced. Recently, there has been a growing body of literature on augmented reality (AR) in adult neurosurgery. In contrast, the use of AR in pediatric neurosurgery is comparatively less. Nonetheless, we postulate that AR systems will be helpful for understanding spatial relationships of complex GSBT anatomy for preoperative planning in a timely fashion. This study describes our experience in trialing AR as a potential tool for three cases of pediatric GSBTs. Overall, the AR platform offers our neurosurgical team excellent visuospatial insights for preoperative decision-making. However, we observe that substantial time is required to set up the AR system prior to each clinical case discussion by the neurosurgical team. In congruency with existing literature, our preliminary results report that there are still obstacles that need to be addressed before the technology can be seamlessly implemented into the clinical workflow for these time-sensitive childhood brain tumors. To our knowledge, this is the first study to report the potential use of AR for complex pediatric GSBT cases.

A Systematic Review of Automated Segmentation Methods and Public Datasets for the Lung and its Lobes and Findings on Computed Tomography Images

OBJECTIVES

Automated computational segmentation of the lung and its lobes and findings in X-Ray based computed tomography (CT) images is a challenging problem with important applications, including medical research, surgical planning, and diagnostic decision support. With the increase in large imaging cohorts and the need for fast and robust evaluation of normal and abnormal lungs and their lobes, several authors have proposed automated methods for lung assessment on CT images. In this paper we intend to provide a comprehensive summarization of these methods.

METHODS

We used a systematic approach to perform an extensive review of automated lung segmentation methods. We chose Scopus, PubMed, and Scopus to conduct our review and included methods that perform segmentation of the lung parenchyma, lobes or internal disease related findings. The review was not limited by date, but rather by only including methods providing quantitative evaluation.

RESULTS

We organized and classified all 234 included articles into various categories according to methodological similarities among them. We provide summarizations of quantitative evaluations, public datasets, evaluation metrics, and overall statistics indicating recent research directions of the field.

CONCLUSIONS

We noted the rise of data-driven models in the last decade, especially due to the deep learning trend, increasing the demand for high-quality data annotation. This has instigated an increase of semi-supervised and uncertainty guided works that try to be less dependent on human annotation. In addition, the question of how to evaluate the robustness of data-driven methods remains open, given that evaluations derived from specific datasets are not general.

Evaluation of a Dedicated Software "Elements™ Spine SRS, Brainlab®" for Target Volume Definition in the Treatment of Spinal Bone Metastases With Stereotactic Body Radiotherapy

Introduction

Stereotactic body radiotherapy (SBRT) is a treatment option for spine metastases. The International Spine Radiosurgery Consortium (ISRC) has published consensus guidelines for target delineation in spine SBRT. A new software called Elements™ Spine SRS by Brainlab® that includes the module Elements SmartBrush Spine (v3.0, Munich, Germany) has been developed specifically for SBRT treatment of spine metastases, and the latter provides the ability to perform semiautomatic clinical target volume (CTV) generation based on gross tumor volume (GTV) localization and guidelines. The aims of our study were to evaluate this software by studying differences in volumes between semiautomatic CTV contours compared to manual contouring performed by an expert radiation oncologist and to determine the dosimetric impact of these differences on treatment plans.

Methods

A total of 35 volumes ("Expert GTV" and "Expert CTV") from 30 patients were defined by a single expert. A semiautomatic definition of these 35 CTVs based on the location of "Expert GTV" and following ISRC guidelines was also performed in Elements SmartBrush Spine ("Brainlab CTV"). The spatial overlap between "Brainlab" and "Expert" CTVs was calculated using the Dice similarity coefficient (DSC). We considered a threshold of 0.80 or above to indicate that Elements SmartBrush Spine performed very well with adequate contours for clinical use. Two dosimetric treatment plans, each corresponding to a specific planning target volume (PTV; Expert PTV, Brainlab PTV), were created for 11 patients.

Results

We showed that "Brainlab CTV" and "Expert CTV" mean volumes were 29.8 ± 16.1 and 28.7 ± 15.7 cm3, respectively (p = 0.23). We also showed that the mean DSC for semiautomatic contouring relative to expert manual contouring was 0.85 ± 0.08 and less than 0.80 in five cases. For metastases involving the vertebral body only (n = 13,37%), the mean DSC was 0.90 ± 0.03, and for ones involving other or several vertebral regions (n = 22.63%), the mean DSC was 0.81 ± 0.08 (p < 0.001). The comparison of dosimetric treatment plans was performed for equivalent PTV coverage. There were no differences between doses received by organs at risk (spinal cord and esophagus) for Expert and Brainlab PTVs, respectively.

Conclusion

The results showed that the semiautomatic method had quite good accuracy and can be used in clinical routine even for complex lesions.

Clinically Oriented Target Contour Evaluation Using Geometric and Dosimetric Indices Based on Simple Geometric Transformations

Purpose:

In radiotherapy, geometric indices are often used to evaluate the accuracy of contouring. However, the ability of geometric indices to identify the error of contouring results is limited primarily because they do not consider the clinical background. The purpose of this study is to investigate the relationship between geometric and clinical dosimetric indices.

Methods:

Four different types of targets were selected (C-shaped target, oropharyngeal cancer, metastatic spine cancer, and prostate cancer), and the translation, scaling, rotation, and sine function transformation were performed with the software Python to introduce systematic and random errors. The transformed contours were regarded as reference contours. Dosimetric indices were obtained from the original dose distribution of the radiotherapy plan. The correlations between geometric and dosimetric indices were quantified by linear regression.

Results:

The correlations between the geometric and dosimetric indices were inconsistent. For systematic errors, and with the exception of the sine function transformation (R²: 0.023-0.04, P > 0.05), the geometric transformations of the C-shaped target were correlated with the D98% and D_mean (R²: 0.689-0.988), 80% of which were P < 0.001. For the random errors, the correlations obtained by the all targets were R² > 0.384, P < 0.05. The Wilcoxon signed-rank test was used to compare the spatial direction resolution capability of geometric indices in different directions of the C-shaped target (with systematic errors), and the results showed only the volumetric geometric indices with P < 0.05.

Conclusions:

Clinically, an assessment of the contour accuracy of the region-of-interest is not feasible based on geometric indices alone. Dosimetric indices should be added to the evaluations of the accuracy of the delineation results, which can be helpful for explaining the clinical dose response relationship of delineation more comprehensively and accurately.

Training and Validation of Deep Learning-Based Auto-Segmentation Models for Lung Stereotactic Ablative Radiotherapy Using Retrospective Radiotherapy Planning Contours

PURPOSE

Deep learning-based auto-segmented contour (DC) models require high quality data for their development, and previous studies have typically used prospectively produced contours, which can be resource intensive and time consuming to obtain. The aim of this study was to investigate the feasibility of using retrospective peer-reviewed radiotherapy planning contours in the training and evaluation of DC models for lung stereotactic ablative radiotherapy (SABR).

METHODS

Using commercial deep learning-based auto-segmentation software, DC models for lung SABR organs at risk (OAR) and gross tumor volume (GTV) were trained using a deep convolutional neural network and a median of 105 contours per structure model obtained from 160 publicly available CT scans and 50 peer-reviewed SABR planning 4D-CT scans from center A. DCs were generated for 50 additional planning CT scans from center A and 50 from center B, and compared with the clinical contours (CC) using the Dice Similarity Coefficient (DSC) and 95% Hausdorff distance (HD).

RESULTS

Comparing DCs to CCs, the mean DSC and 95% HD were 0.93 and 2.85mm for aorta, 0.81 and 3.32mm for esophagus, 0.95 and 5.09mm for heart, 0.98 and 2.99mm for bilateral lung, 0.52 and 7.08mm for bilateral brachial plexus, 0.82 and 4.23mm for proximal bronchial tree, 0.90 and 1.62mm for spinal cord, 0.91 and 2.27mm for trachea, and 0.71 and 5.23mm for GTV. DC to CC comparisons of center A and center B were similar for all OAR structures.

CONCLUSIONS

The DCs developed with retrospective peer-reviewed treatment contours approximated CCs for the majority of OARs, including on an external dataset. DCs for structures with more variability tended to be less accurate and likely require using a larger number of training cases or novel training approaches to improve performance. Developing DC models from existing radiotherapy planning contours appears feasible and warrants further clinical workflow testing.

Comparison between atlas and convolutional neural network based automatic segmentation of multiple organs at risk in non-small cell lung cancer

Delineation of organs at risk (OARs) is important but time consuming for radiotherapy planning. Automatic segmentation of OARs based on convolutional neural network (CNN) has been established for lung cancer patients at our institution. The aim of this study is to compare automatic segmentation based on CNN (AS-CNN) with automatic segmentation based on atlas (AS-Atlas) in terms of the efficiency and accuracy of OARs contouring.The OARs, including the lungs, esophagus, heart, liver, and spinal cord, of 19 non-small cell lung cancer patients were delineated using three methods: AS-CNN, AS-Atlas in the Pinnacle-software, and manual delineation (MD) by a senior radiation oncologist. MD was used as the ground-truth reference, and the segmentation efficiency was evaluated by the time spent per patient. The accuracy was evaluated using the Mean surface distance (MSD) and Dice similarity coefficient (DSC). The paired t-test or Wilcoxon signed-rank test was used to compare these indexes between the 2 automatic segmentation models.In the 19 testing cases, both AS-CNN and AS-Atlas saved substantial time compared with MD. AS-CNN was more efficient than AS-Atlas (1.6 min vs 2.4 min, P < .001). In terms of the accuracy, AS-CNN performed well in the esophagus, with a DSC of 73.2%. AS-CNN was better than AS-Atlas in segmenting the left lung (DSC: 94.8% vs 93.2%, P = .01; MSD: 1.10 cm vs 1.73 cm, P < .001) and heart (DSC: 89.3% vs 85.8%, P = .05; MSD: 1.65 cm vs 3.66 cm, P < .001). Furthermore, AS-CNN exhibited superior performance in segmenting the liver (DSC: 93.7% vs 93.6%, P = .81; MSD: 2.03 cm VS 2.11 cm, P = .66). The results obtained from AS-CNN and AS-Atlas were similar in segmenting the right lung. However, the performance of AS-CNN in the spinal cord was inferior to that of AS-Atlas (DSC: 82.1% vs 86.8%, P = .01; MSD: 0.87 cm vs 0.66 cm, P = .01).Our study demonstrated that AS-CNN significantly reduced the contouring time and outperformed AS-Atlas in most cases. AS-CNN can potentially be used for OARs segmentation in patients with pathological N2 (pN2) non-small cell lung cancer.