AAR-RT - A system for auto-contouring organs at risk on CT images for radiation therapy planning: Principles, design, and large-scale evaluation on head-and-neck and thoracic cancer cases

Deep learning for autosegmentation for radiotherapy treatment planning: State-of-the-art and novel perspectives

The rapid development of artificial intelligence (AI) has gained importance, with many tools already entering our daily lives. The medical field of radiation oncology is also subject to this development, with AI entering all steps of the patient journey. In this review article, we summarize contemporary AI techniques and explore the clinical applications of AI-based automated segmentation models in radiotherapy planning, focusing on delineation of organs at risk (OARs), the gross tumor volume (GTV), and the clinical target volume (CTV). Emphasizing the need for precise and individualized plans, we review various commercial and freeware segmentation tools and also state-of-the-art approaches. Through our own findings and based on the literature, we demonstrate improved efficiency and consistency as well as time savings in different clinical scenarios. Despite challenges in clinical implementation such as domain shifts, the potential benefits for personalized treatment planning are substantial. The integration of mathematical tumor growth models and AI-based tumor detection further enhances the possibilities for refining target volumes. As advancements continue, the prospect of one-stop-shop segmentation and radiotherapy planning represents an exciting frontier in radiotherapy, potentially enabling fast treatment with enhanced precision and individualization.

Automated pancreatic segmentation and fat fraction evaluation based on a self-supervised transfer learning network

Accurate segmentation of the pancreas from abdominal computed tomography (CT) images is challenging but essential for the diagnosis and treatment of pancreatic disorders such as tumours and diabetes. In this study, a dataset with 229 sets of high-resolution CT images was generated and annotated. We proposed a novel 3D segmentation model named nnTransfer (nonisomorphic transfer learning) net, which employs generative model structure for self-supervision to facilitate the network's learning of image attributes from unlabelled data. The effectiveness for pancreas segmentation of nnTransfer was assessed using the Hausdorff distance (HD) and Dice similarity coefficient (DSC) on the dataset. Additionally, a histogram analysis with local thresholding was used to achieve automated whole-volume measurement of pancreatic fat (fat volume fraction, FVF). The proposed technique performed admirably on the dataset, with DSC: 0.937 ± 0.019 and HD: 2.655 ± 1.479. The mean pancreas volume and FVF of the pancreas were 91.95 ± 23.90 cm3 and 12.67 % ± 9.84 %, respectively. The nnTransfer functioned flawlessly and autonomously, facilitating the use of the FVF to evaluate pancreatic disease, particularly in patients with diabetes.

vOARiability: Interobserver and intermodality variability analysis in OAR contouring from head and neck CT and MR images

BACKGROUND

Accurate and consistent contouring of organs-at-risk (OARs) from medical images is a key step of radiotherapy (RT) cancer treatment planning. Most contouring approaches rely on computed tomography (CT) images, but the integration of complementary magnetic resonance (MR) modality is highly recommended, especially from the perspective of OAR contouring, synthetic CT and MR image generation for MR-only RT, and MR-guided RT. Although MR has been recognized as valuable for contouring OARs in the head and neck (HaN) region, the accuracy and consistency of the resulting contours have not been yet objectively evaluated.

PURPOSE

To analyze the interobserver and intermodality variability in contouring OARs in the HaN region, performed by observers with different level of experience from CT and MR images of the same patients.

METHODS

In the final cohort of 27 CT and MR images of the same patients, contours of up to 31 OARs were obtained by a radiation oncology resident (junior observer, JO) and a board-certified radiation oncologist (senior observer, SO). The resulting contours were then evaluated in terms of interobserver variability, characterized as the agreement among different observers (JO and SO) when contouring OARs in a selected modality (CT or MR), and intermodality variability, characterized as the agreement among different modalities (CT and MR) when OARs were contoured by a selected observer (JO or SO), both by the Dice coefficient (DC) and 95-percentile Hausdorff distance (HD95 $_{95}$ ).

RESULTS

The mean (±standard deviation) interobserver variability was 69.0 ± 20.2% and 5.1 ± 4.1 mm, while the mean intermodality variability was 61.6 ± 19.0% and 6.1 ± 4.3 mm in terms of DC and HD95 $_{95}$ , respectively, across all OARs. Statistically significant differences were only found for specific OARs. The performed MR to CT image registration resulted in a mean target registration error of 1.7 ± 0.5 mm, which was considered as valid for the analysis of intermodality variability.

CONCLUSIONS

The contouring variability was, in general, similar for both image modalities, and experience did not considerably affect the contouring performance. However, the results indicate that an OAR is difficult to contour regardless of whether it is contoured in the CT or MR image, and that observer experience may be an important factor for OARs that are deemed difficult to contour. Several of the differences in the resulting variability can be also attributed to adherence to guidelines, especially for OARs with poor visibility or without distinctive boundaries in either CT or MR images. Although considerable contouring differences were observed for specific OARs, it can be concluded that almost all OARs can be contoured with a similar degree of variability in either the CT or MR modality, which works in favor of MR images from the perspective of MR-only and MR-guided RT.

Optimal strategies for modeling anatomy in a hybrid intelligence framework for auto-segmentation of organs

Organ segmentation is a fundamental requirement in medical image analysis. Many methods have been proposed over the past 6 decades for segmentation. A unique feature of medical images is the anatomical information hidden within the image itself. To bring natural intelligence (NI) in the form of anatomical information accumulated over centuries into deep learning (DL) AI methods effectively, we have recently introduced the idea of hybrid intelligence (HI) that combines NI and AI and a system based on HI to perform medical image segmentation. This HI system has shown remarkable robustness to image artifacts, pathology, deformations, etc. in segmenting organs in the Thorax body region in a multicenter clinical study. The HI system utilizes an anatomy modeling strategy to encode NI and to identify a rough container region in the shape of each object via a non-DL-based approach so that DL training and execution are applied only to the fuzzy container region. In this paper, we introduce several advances related to modeling of the NI component so that it becomes substantially more efficient computationally, and at the same time, is well integrated with the DL portion (AI component) of the system. We demonstrate a 9-40 fold computational improvement in the auto-segmentation task for radiation therapy (RT) planning via clinical studies obtained from 4 different RT centers, while retaining state-of-the-art accuracy of the previous system in segmenting 11 objects in the Thorax body region.

Educative Impact of Automatic Delineation Applied to Head and Neck Cancer Patients on Radiation Oncology Residents

To evaluate the educational impact on radiation oncology residents in training when introducing an automatic segmentation software in head and neck cancer patients regarding organs at risk (OARs) and prophylactic cervical lymph node level (LNL) volumes. Two cases treated by exclusive intensity-modulated radiotherapy were delineated by an expert radiation oncologist and were considered as reference. Then, these cases were delineated by residents divided into two groups: group 1 (control group), experienced residents delineating manually, group 2 (experimental group), young residents on their first rotation trained with automatic delineation, delineating manually first (M -) and then after using the automatic system (M +). The delineation accuracy was assessed using the Overlap Volume (OV). Regarding the OARs, mean OV was 0.62 (SD = 0.05) for group 1, 0.56 (SD = 0.04) for group 2 M - , and 0.61 (SD = 0.03) for group 2 M + . Mean OV was higher in group 1 compared to group 2 M - (p = 0.01). There was no OV difference between group 1 and group 2 M + (p = 0.67). Mean OV was higher in the group 2 M + compared to group 2 M - (p < 0.003). Regarding LNL, mean OV was 0.53 (SD = 0.06) in group 1, 0.54 (SD = 0.03) in group 2 M - , and 0.58 (SD = 0.04) in group 2 M + . Mean OV was higher in group 2 M + for 11 of the 12 analysed structures compared to group 2 M - (p = 0.016). Prior use of the automatic delineation software reduced the average contouring time per case by 34 to 40%. Prior use of atlas-based automatic segmentation reduces the delineation duration, and provides reliable OARs and LNL delineations.

Integration of artificial intelligence in lung cancer: Rise of the machine

The goal of oncology is to provide the longest possible survival outcomes with the therapeutics that are currently available without sacrificing patients' quality of life. In lung cancer, several data points over a patient's diagnostic and treatment course are relevant to optimizing outcomes in the form of precision medicine, and artificial intelligence (AI) provides the opportunity to use available data from molecular information to radiomics, in combination with patient and tumor characteristics, to help clinicians provide individualized care. In doing so, AI can help create models to identify cancer early in diagnosis and deliver tailored therapy on the basis of available information, both at the time of diagnosis and in real time as they are undergoing treatment. The purpose of this review is to summarize the current literature in AI specific to lung cancer and how it applies to the multidisciplinary team taking care of these complex patients.

Auto-segmentation of thoracic brachial plexuses for radiation therapy planning

Recently, deep learning networks have achieved considerable success in segmenting organs in medical images. Several methods have used volumetric information with deep networks to achieve segmentation accuracy. However, these networks suffer from interference, risk of overfitting, and low accuracy as a result of artifacts, in the case of very challenging objects like the brachial plexuses. In this paper, to address these issues, we synergize the strengths of high-level human knowledge (i.e., natural intelligence (NI)) with deep learning (i.e., artificial intelligence (AI)) for recognition and delineation of the thoracic brachial plexuses (BPs) in computed tomography (CT) images. We formulate an anatomy-guided deep learning hybrid intelligence approach for segmenting thoracic right and left brachial plexuses consisting of 2 key stages. In the first stage (AAR-R), objects are recognized based on a previously created fuzzy anatomy model of the body region with its key organs relevant for the task at hand wherein high-level human anatomic knowledge is precisely codified. The second stage (DL-D) uses information from AAR-R to limit the search region to just where each object is most likely to reside and performs encoder-decoder delineation in slices. The proposed method is tested on a dataset that consists of 125 images of the thorax acquired for radiation therapy planning of tumors in the thorax and achieves a Dice coefficient of 0.659.

Combining natural and artificial intelligence for robust automatic anatomy segmentation: Application in neck and thorax auto-contouring

BACKGROUND

Automatic segmentation of 3D objects in computed tomography (CT) is challenging. Current methods, based mainly on artificial intelligence (AI) and end-to-end deep learning (DL) networks, are weak in garnering high-level anatomic information, which leads to compromised efficiency and robustness. This can be overcome by incorporating natural intelligence (NI) into AI methods via computational models of human anatomic knowledge.

PURPOSE

We formulate a hybrid intelligence (HI) approach that integrates the complementary strengths of NI and AI for organ segmentation in CT images and illustrate performance in the application of radiation therapy (RT) planning via multisite clinical evaluation.

METHODS

The system employs five modules: (i) body region recognition, which automatically trims a given image to a precisely defined target body region; (ii) NI-based automatic anatomy recognition object recognition (AAR-R), which performs object recognition in the trimmed image without DL and outputs a localized fuzzy model for each object; (iii) DL-based recognition (DL-R), which refines the coarse recognition results of AAR-R and outputs a stack of 2D bounding boxes (BBs) for each object; (iv) model morphing (MM), which deforms the AAR-R fuzzy model of each object guided by the BBs output by DL-R; and (v) DL-based delineation (DL-D), which employs the object containment information provided by MM to delineate each object. NI from (ii), AI from (i), (iii), and (v), and their combination from (iv) facilitate the HI system.

RESULTS

The HI system was tested on 26 organs in neck and thorax body regions on CT images obtained prospectively from 464 patients in a study involving four RT centers. Data sets from one separate independent institution involving 125 patients were employed in training/model building for each of the two body regions, whereas 104 and 110 data sets from the 4 RT centers were utilized for testing on neck and thorax, respectively. In the testing data sets, 83% of the images had limitations such as streak artifacts, poor contrast, shape distortion, pathology, or implants. The contours output by the HI system were compared to contours drawn in clinical practice at the four RT centers by utilizing an independently established ground-truth set of contours as reference. Three sets of measures were employed: accuracy via Dice coefficient (DC) and Hausdorff boundary distance (HD), subjective clinical acceptability via a blinded reader study, and efficiency by measuring human time saved in contouring by the HI system. Overall, the HI system achieved a mean DC of 0.78 and 0.87 and a mean HD of 2.22 and 4.53 mm for neck and thorax, respectively. It significantly outperformed clinical contouring in accuracy and saved overall 70% of human time over clinical contouring time, whereas acceptability scores varied significantly from site to site for both auto-contours and clinically drawn contours.

CONCLUSIONS

The HI system is observed to behave like an expert human in robustness in the contouring task but vastly more efficiently. It seems to use NI help where image information alone will not suffice to decide, first for the correct localization of the object and then for the precise delineation of the boundary.

A survey of medical dosimetrists' perceptions of efficiency and consistency of auto-contouring software

Although auto-contouring methods were created to reduce the workload for the radiation oncology team, concern lies in whether auto-contouring can improve efficiency regarding generated contours of a treatment plan. Researchers have measured differences between auto-contouring algorithms and manual contour methods specific to the contouring of organs at risk (OAR). The problem lies in the paucity of literature specific to perceptions of auto-contouring and the impact on workflow efficiency. The purpose of this study was to measure medical dosimetrists' perceptions of how auto-contouring software impacts the treatment planning process. To measure perceptions, researchers surveyed medical dosimetrists about their perspectives on consistency and efficiency of auto-contouring during treatment planning. A (Qualtrics, Provo, UT) survey was created based on the 2 research questions in this study. The survey was distributed through email to 2598 full members of the American Association of Medical Dosimetrists (AAMD) who were certified by the MDCB; mostly medical dosimetrists but also included a small group of medical physicists. The email open rate was 39% (1024/2598) but the response rate for those who read the email was only 8.4% (86/1024). Of the survey respondents, 67% (59/86) used auto-contouring software; thus, eligible to complete the remainder of the survey. Majority of participants agreed that auto-contouring software decreases time spent contouring per patient; however, most agreed that manual contouring is more efficient. Therefore, it was inferred that a combination of both auto and manual contouring have an impact on workload efficiency.

SOMA: Subject-, object-, and modality-adapted precision atlas approach for automatic anatomy recognition and delineation in medical images

PURPOSE

In the multi-atlas segmentation (MAS) method, a large enough atlas set, which can cover the complete spectrum of the whole population pattern of the target object will benefit the segmentation quality. However, the difficulty in obtaining and generating such a large set of atlases and the computational burden required in the segmentation procedure make this approach impractical. In this paper, we propose a method called SOMA to select subject-, object-, and modality-adapted precision atlases for automatic anatomy recognition in medical images with pathology, following the idea that different regions of the target object in a novel image can be recognized by different atlases with regionally best similarity, so that effective atlases have no need to be globally similar to the target subject and also have no need to be overall similar to the target object.

METHODS

The SOMA method consists of three main components: atlas building, object recognition, and object delineation. Considering the computational complexity, we utilize an all-to-template strategy to align all images to the same image space belonging to the root image determined by the minimum spanning tree (MST) strategy among a subset of radiologically near-normal images. The object recognition process is composed of two stages: rough recognition and refined recognition. In rough recognition, subimage matching is conducted between the test image and each image of the whole atlas set, and only the atlas corresponding to the best-matched subimage contributes to the recognition map regionally. The frequency of best match for each atlas is recorded by a counter, and the atlases with the highest frequencies are selected as the precision atlases. In refined recognition, only the precision atlases are examined, and the subimage matching is conducted in a nonlocal manner of searching to further increase the accuracy of boundary matching. Delineation is based on a U-net-based deep learning network, where the original gray scale image together with the fuzzy map from refined recognition compose a two-channel input to the network, and the output is a segmentation map of the target object.

RESULTS

Experiments are conducted on computed tomography (CT) images with different qualities in two body regions - head and neck (H&N) and thorax, from 298 subjects with nine objects and 241 subjects with six objects, respectively. Most objects achieve a localization error within two voxels after refined recognition, with marked improvement in localization accuracy from rough to refined recognition of 0.6-3 mm in H&N and 0.8-4.9 mm in thorax, and also in delineation accuracy (Dice coefficient) from refined recognition to delineation of 0.01-0.11 in H&N and 0.01-0.18 in thorax.

CONCLUSIONS

The SOMA method shows high accuracy and robustness in anatomy recognition and delineation. The improvements from rough to refined recognition and further to delineation, as well as immunity of recognition accuracy to varying image and object qualities, demonstrate the core principles of SOMA where segmentation accuracy increases with precision atlases and gradually refined object matching.

Automatic upper airway segmentation in static and dynamic MRI via anatomy-guided convolutional neural networks

PURPOSE

Upper airway segmentation on MR images is a prerequisite step for quantitatively studying the anatomical structure and function of the upper airway and surrounding tissues. However, the complex variability of intensity and shape of anatomical structures and different modes of image acquisition commonly used in this application makes automatic upper airway segmentation challenging. In this paper, we develop and test a comprehensive deep learning-based segmentation system for use on MR images to address this problem.

MATERIALS AND METHODS

In our study, both static and dynamic MRI data sets are utilized, including 58 axial static 3D MRI studies, 22 mid-retropalatal dynamic 2D MRI studies, 21 mid-retroglossal dynamic 2D MRI studies, 36 mid-sagittal dynamic 2D MRI studies, and 23 isotropic dynamic 3D MRI studies, involving a total of 160 subjects and over 20 000 MRI slices. Samples of static and 2D dynamic MRI data sets were randomly divided into training, validation, and test sets by an approximate ratio of 5:2:3. Considering that the variability of annotation data among 3D dynamic MRIs was greater than for other MRI data sets, we increased the ratio of training data for these data to improve the robustness of the model. We designed a unified framework consisting of the following procedures. For static MRI, a generalized region-of-interest (GROI) strategy is applied to localize the partitions of nasal cavity and other portions of upper airway in axial data sets as two separate subobjects. Subsequently, the two subobjects are segmented by two separate 2D U-Nets. The two segmentation results are combined as the whole upper airway structure. The GROI strategy is also applied to other MRI modes. To minimize false-positive and false-negative rates in the segmentation results, we employed a novel loss function based explicitly on these rates to train the segmentation networks. An inter-reader study is conducted to test the performance of our system in comparison to human variability in ground truth (GT) segmentation of these challenging structures.

RESULTS

The proposed approach yielded mean Dice coefficients of 0.84±0.03, 0.89±0.13, 0.84±0.07, and 0.86±0.05 for static 3D MRI, mid-retropalatal/mid-retroglossal 2D dynamic MRI, mid-sagittal 2D dynamic MRI, and isotropic dynamic 3D MRI, respectively. The quantitative results show excellent agreement with manual delineation results. The inter-reader study results demonstrate that the segmentation performance of our approach is statistically indistinguishable from manual segmentations considering the inter-reader variability in GT.

CONCLUSIONS

The proposed method can be utilized for routine upper airway segmentation from static and dynamic MR images with high accuracy and efficiency. The proposed approach has the potential to be employed in other dynamic MRI-related applications, such as lung or heart segmentation.

Clinically Applicable Segmentation of Head and Neck Anatomy for Radiotherapy: Deep Learning Algorithm Development and Validation Study

BACKGROUND

Over half a million individuals are diagnosed with head and neck cancer each year globally. Radiotherapy is an important curative treatment for this disease, but it requires manual time to delineate radiosensitive organs at risk. This planning process can delay treatment while also introducing interoperator variability, resulting in downstream radiation dose differences. Although auto-segmentation algorithms offer a potentially time-saving solution, the challenges in defining, quantifying, and achieving expert performance remain.

OBJECTIVE

Adopting a deep learning approach, we aim to demonstrate a 3D U-Net architecture that achieves expert-level performance in delineating 21 distinct head and neck organs at risk commonly segmented in clinical practice.

METHODS

The model was trained on a data set of 663 deidentified computed tomography scans acquired in routine clinical practice and with both segmentations taken from clinical practice and segmentations created by experienced radiographers as part of this research, all in accordance with consensus organ at risk definitions.

RESULTS

We demonstrated the model's clinical applicability by assessing its performance on a test set of 21 computed tomography scans from clinical practice, each with 21 organs at risk segmented by 2 independent experts. We also introduced surface Dice similarity coefficient, a new metric for the comparison of organ delineation, to quantify the deviation between organ at risk surface contours rather than volumes, better reflecting the clinical task of correcting errors in automated organ segmentations. The model's generalizability was then demonstrated on 2 distinct open-source data sets, reflecting different centers and countries to model training.

CONCLUSIONS

Deep learning is an effective and clinically applicable technique for the segmentation of the head and neck anatomy for radiotherapy. With appropriate validation studies and regulatory approvals, this system could improve the efficiency, consistency, and safety of radiotherapy pathways.

Automatic Segmentation of Mandible from Conventional Methods to Deep Learning-A Review

Medical imaging techniques, such as (cone beam) computed tomography and magnetic resonance imaging, have proven to be a valuable component for oral and maxillofacial surgery (OMFS). Accurate segmentation of the mandible from head and neck (H&N) scans is an important step in order to build a personalized 3D digital mandible model for 3D printing and treatment planning of OMFS. Segmented mandible structures are used to effectively visualize the mandible volumes and to evaluate particular mandible properties quantitatively. However, mandible segmentation is always challenging for both clinicians and researchers, due to complex structures and higher attenuation materials, such as teeth (filling) or metal implants that easily lead to high noise and strong artifacts during scanning. Moreover, the size and shape of the mandible vary to a large extent between individuals. Therefore, mandible segmentation is a tedious and time-consuming task and requires adequate training to be performed properly. With the advancement of computer vision approaches, researchers have developed several algorithms to automatically segment the mandible during the last two decades. The objective of this review was to present the available fully (semi)automatic segmentation methods of the mandible published in different scientific articles. This review provides a vivid description of the scientific advancements to clinicians and researchers in this field to help develop novel automatic methods for clinical applications.

Segmentation evaluation with sparse ground truth data: Simulating true segmentations as perfect/imperfect as those generated by humans

Fully annotated data sets play important roles in medical image segmentation and evaluation. Expense and imprecision are the two main issues in generating ground truth (GT) segmentations. In this paper, in an attempt to overcome these two issues jointly, we propose a method, named SparseGT, which exploit variability among human segmenters to maximally save manual workload in GT generation for evaluating actual segmentations by algorithms. Pseudo ground truth (p-GT) segmentations are created by only a small fraction of workload and with human-level perfection/imperfection, and they can be used in practice as a substitute for fully manual GT in evaluating segmentation algorithms at the same precision. p-GT segmentations are generated by first selecting slices sparsely, where manual contouring is conducted only on these sparse slices, and subsequently filling segmentations on other slices automatically. By creating p-GT with different levels of sparseness, we determine the largest workload reduction achievable for each considered object, where the variability of the generated p-GT is statistically indistinguishable from inter-segmenter differences in full manual GT segmentations for that object. Furthermore, we investigate the segmentation evaluation errors introduced by variability in manual GT by applying p-GT in evaluation of actual segmentations by an algorithm. Experiments are conducted on ∼500 computed tomography (CT) studies involving six objects in two body regions, Head & Neck and Thorax, where optimal sparseness and corresponding evaluation errors are determined for each object and each strategy. Our results indicate that creating p-GT by the concatenated strategy of uniformly selecting sparse slices and filling segmentations via deep-learning (DL) network show highest manual workload reduction by ∼80-96% without sacrificing evaluation accuracy compared to fully manual GT. Nevertheless, other strategies also have obvious contributions in different situations. A non-uniform strategy for slice selection shows its advantage for objects with irregular shape change from slice to slice. An interpolation strategy for filling segmentations can achieve ∼60-90% of workload reduction in simulating human-level GT without the need of an actual training stage and shows potential in enlarging data sets for training p-GT generation networks. We conclude that not only over 90% reduction in workload is feasible without sacrificing evaluation accuracy but also the suitable strategy and the optimal sparseness level achievable for creating p-GT are object- and application-specific.

Anatomy Recognition in CT Images of Head & Neck Region via Precision Atlases

Multi-atlas segmentation methods will benefit from atlases covering the complete spectrum of population patterns, while the difficulties in generating such large enough datasets and the computation burden required in the segmentation procedure reduce its practicality in clinical application. In this work, we start from a viewpoint that different parts of the target object can be recognized by different atlases and propose a precision atlas selection strategy. By comparing regional similarity between target image and atlases, precision atlases are ranked and selected by the frequency of regional best match, which have no need to be globally similar to the target subject at either image-level or object-level, largely increasing the implicit patterns contained in the atlas set. In the proposed anatomy recognition method, atlas building is first achieved by all-to-template registration, where the minimum spanning tree (MST) strategy is used to select a registration template from a subset of radiologically near-normal images. Then, a two-stage recognition process is conducted: in rough recognition, sub-image level similarity is calculated between the test image and each image of the whole atlas set, and only the atlas with the highest similarity contributes to the recognition map regionally; in refined recognition, the atlases with the highest frequencies of best match are selected as the precision atlases and are utilized to further increase the accuracy of boundary matching. The proposed method is demonstrated on 298 computed tomography (CT) images and 9 organs in the Head & Neck (H&N) body region. Experimental results illustrate that our method is effective for organs with different segmentation challenge and samples with different image quality, where remarkable improvement in boundary interpretation is made by refined recognition and most objects achieve a localization error within 2 voxels.

A slice classification model-facilitated 3D encoder-decoder network for segmenting organs at risk in head and neck cancer

For deep learning networks used to segment organs at risk (OARs) in head and neck (H&N) cancers, the class-imbalance problem between small volume OARs and whole computed tomography (CT) images results in delineation with serious false-positives on irrelevant slices and unnecessary time-consuming calculations. To alleviate this problem, a slice classification model-facilitated 3D encoder-decoder network was developed and validated. In the developed two-step segmentation model, a slice classification model was firstly utilized to classify CT slices into six categories in the craniocaudal direction. Then the target categories for different OARs were pushed to the different 3D encoder-decoder segmentation networks, respectively. All the patients were divided into training (n = 120), validation (n = 30) and testing (n = 20) datasets. The average accuracy of the slice classification model was 95.99%. The Dice similarity coefficient and 95% Hausdorff distance, respectively, for each OAR were as follows: right eye (0.88 ± 0.03 and 1.57 ± 0.92 mm), left eye (0.89 ± 0.03 and 1.35 ± 0.43 mm), right optic nerve (0.72 ± 0.09 and 1.79 ± 1.01 mm), left optic nerve (0.73 ± 0.09 and 1.60 ± 0.71 mm), brainstem (0.87 ± 0.04 and 2.28 ± 0.99 mm), right temporal lobe (0.81 ± 0.12 and 3.28 ± 2.27 mm), left temporal lobe (0.82 ± 0.09 and 3.73 ± 2.08 mm), right temporomandibular joint (0.70 ± 0.13 and 1.79 ± 0.79 mm), left temporomandibular joint (0.70 ± 0.16 and 1.98 ± 1.48 mm), mandible (0.89 ± 0.02 and 1.66 ± 0.51 mm), right parotid (0.77 ± 0.07 and 7.30 ± 4.19 mm) and left parotid (0.71 ± 0.12 and 8.41 ± 4.84 mm). The total segmentation time was 40.13 s. The 3D encoder-decoder network facilitated by the slice classification model demonstrated superior performance in accuracy and efficiency in segmenting OARs in H&N CT images. This may significantly reduce the workload for radiation oncologists.

BRR-Net: A tandem architectural CNN-RNN for automatic body region localization in CT images

PURPOSE

Automatic identification of consistently defined body regions in medical images is vital in many applications. In this paper, we describe a method to automatically demarcate the superior and inferior boundaries for neck, thorax, abdomen, and pelvis body regions in computed tomography (CT) images.

METHODS

For any three-dimensional (3D) CT image I, following precise anatomic definitions, we denote the superior and inferior axial boundary slices of the neck, thorax, abdomen, and pelvis body regions by NS(I), NI(I), TS(I), TI(I), AS(I), AI(I), PS(I), and PI(I), respectively. Of these, by definition, AI(I) = PS(I), and so the problem reduces to demarcating seven body region boundaries. Our method consists of a two-step approach. In the first step, a convolutional neural network (CNN) is trained to classify each axial slice in I into one of nine categories: the seven body region boundaries, plus legs (defined as all axial slices inferior to PI(I)), and the none-of-the-above category. This CNN uses a multichannel approach to exploit the interslice contrast, providing the neural network with additional visual context at the body region boundaries. In the second step, to improve the predictions for body region boundaries that are very subtle and that exhibit low contrast, a recurrent neural network (RNN) is trained on features extracted by CNN, limited to a flexible window about the predictions from the CNN.

RESULTS

The method is evaluated on low-dose CT images from 442 patient scans, divided into training and testing sets with a ratio of 70:30. Using only the CNN, overall absolute localization error for NS(I), NI(I), TS(I), TI(I), AS(I), AI(I), and PI(I) expressed in terms of number of slices (nS) is (mean ± SD): 0.61 ± 0.58, 1.05 ± 1.13, 0.31 ± 0.46, 1.85 ± 1.96, 0.57 ± 2.44, 3.42 ± 3.16, and 0.50 ± 0.50, respectively. Using the RNN to refine the CNN's predictions for select classes improved the accuracy of TI(I) and AI(I) to: 1.35 ± 1.71 and 2.83 ± 2.75, respectively. This model outperforms the results achieved in our previous work by 2.4, 1.7, 3.1, 1.1, and 2 slices, respectively for TS(I), TI(I), AS(I), AI(I) = PS(I), and PI(I) classes with statistical significance. The model trained on low-dose CT images was also tested on diagnostic CT images for NS(I), NI(I), and TS(I) classes; the resulting errors were: 1.48 ± 1.33, 2.56 ± 2.05, and 0.58 ± 0.71, respectively.

CONCLUSIONS

Standardized body region definitions are a prerequisite for effective implementation of quantitative radiology, but the literature is severely lacking in the precise identification of body regions. The method presented in this paper significantly outperforms earlier works by a large margin, and the deviations of our results from ground truth are comparable to variations observed in manual labeling by experts. The solution presented in this work is critical to the adoption and employment of the idea of standardized body regions, and clears the path for development of applications requiring accurate demarcations of body regions. The work is indispensable for automatic anatomy recognition, delineation, and contouring for radiation therapy planning, as it not only automates an essential part of the process, but also removes the dependency on experts for accurately demarcating body regions in a study.

Auto-segmentation of organs at risk for head and neck radiotherapy planning: From atlas-based to deep learning methods

Radiotherapy (RT) is one of the basic treatment modalities for cancer of the head and neck (H&N), which requires a precise spatial description of the target volumes and organs at risk (OARs) to deliver a highly conformal radiation dose to the tumor cells while sparing the healthy tissues. For this purpose, target volumes and OARs have to be delineated and segmented from medical images. As manual delineation is a tedious and time-consuming task subjected to intra/interobserver variability, computerized auto-segmentation has been developed as an alternative. The field of medical imaging and RT planning has experienced an increased interest in the past decade, with new emerging trends that shifted the field of H&N OAR auto-segmentation from atlas-based to deep learning-based approaches. In this review, we systematically analyzed 78 relevant publications on auto-segmentation of OARs in the H&N region from 2008 to date, and provided critical discussions and recommendations from various perspectives: image modality - both computed tomography and magnetic resonance image modalities are being exploited, but the potential of the latter should be explored more in the future; OAR - the spinal cord, brainstem, and major salivary glands are the most studied OARs, but additional experiments should be conducted for several less studied soft tissue structures; image database - several image databases with the corresponding ground truth are currently available for methodology evaluation, but should be augmented with data from multiple observers and multiple institutions; methodology - current methods have shifted from atlas-based to deep learning auto-segmentation, which is expected to become even more sophisticated; ground truth - delineation guidelines should be followed and participation of multiple experts from multiple institutions is recommended; performance metrics - the Dice coefficient as the standard volumetric overlap metrics should be accompanied with at least one distance metrics, and combined with clinical acceptability scores and risk assessments; segmentation performance - the best performing methods achieve clinically acceptable auto-segmentation for several OARs, however, the dosimetric impact should be also studied to provide clinically relevant endpoints for RT planning.

AAR-LN-DQ: Automatic anatomy recognition based disease quantification in thoracic lymph node zones via FDG PET/CT images without Nodal Delineation

PURPOSE

The derivation of quantitative information from medical images in a practical manner is essential for quantitative radiology (QR) to become a clinical reality, but still faces a major hurdle because of image segmentation challenges. With the goal of performing disease quantification in lymph node (LN) stations without explicit nodal delineation, this paper presents a novel approach for disease quantification (DQ) by automatic recognition of LN zones and detection of malignant lymph nodes within thoracic LN zones via positron emission tomography/computed tomography (PET/CT) images. Named AAR-LN-DQ, this approach decouples DQ methods from explicit nodal segmentation via an LN recognition strategy involving a novel globular filter and a deep neural network called SegNet.

METHOD

The methodology consists of four main steps: (a) Building lymph node zone models by automatic anatomy recognition (AAR) method. It incorporates novel aspects of model building that relate to finding an optimal hierarchy for organs and lymph node zones in the thorax. (b) Recognizing lymph node zones by the built lymph node models. (c) Detecting pathologic LNs in the recognized zones by using a novel globular filter (g-filter) and a multi-level support vector machine (SVM) classifier. Here, we make use of the general globular shape of LNs to first localize them and then use a multi-level SVM classifier to identify pathologic LNs from among the LNs localized by the g-filter. Alternatively, we designed a deep neural network called SegNet which is trained to directly recognize pathologic nodes within AAR localized LN zones. (d) Disease quantification based on identified pathologic LNs within localized zones. A fuzzy disease map is devised to express the degree of disease burden at each voxel within the identified LNs to simultaneously handle several uncertain phenomena such as PET partial volume effects, uncertainty in localization of LNs, and gradation of disease content at the voxel level. We focused on the task of disease quantification in patients with lymphoma based on PET/CT acquisitions and devised a method of evaluation. Model building was carried out using 42 near-normal patient datasets via contrast-enhanced CT examinations of their thorax. PET/CT datasets from an additional 63 lymphoma patients were utilized for evaluating the AAR-LN-DQ methodology. We assess the accuracy of the three main processes involved in AAR-LN-DQ via fivefold cross validation: lymph node zone recognition, abnormal lymph node localization, and disease quantification.

RESULTS

The recognition and scale error for LN zones were 12.28 mm ± 1.99 and 0.94 ± 0.02, respectively, on normal CT datasets. On abnormal PET/CT datasets, the sensitivity and specificity of pathologic LN recognition were 84.1% ± 0.115 and 98.5% ± 0.003, respectively, for the g-filter-SVM strategy, and 91.3% ± 0.110 and 96.1% ± 0.016, respectively, for the SegNet method. Finally, the mean absolute percent errors for disease quantification of the recognized abnormal LNs were 8% ± 0.09 and 14% ± 0.10 for the g-filter-SVM method and the best SegNet strategy, respectively.

CONCLUSIONS

Accurate disease quantification on PET/CT images without performing explicit delineation of lymph nodes is feasible following lymph node zone and pathologic LN localization. It is very useful to perform LN zone recognition by AAR as this step can cover most (95.8%) of the abnormal LNs and drastically reduce the regions to search for abnormal LNs. This also improves the specificity of deep networks such as SegNet significantly. It is possible to utilize general shape information about LNs such as their globular nature via g-filter and to arrive at high recognition rates for abnormal LNs in conjunction with a traditional classifier such as SVM. Finally, the disease map concept is effective for estimating disease burden, irrespective of how the LNs are identified, to handle various uncertainties without having to address them explicitly one by one.

Obtaining the potential number of object models/atlases needed in medical image analysis

Medical image processing and analysis operations, particularly segmentation, can benefit a great deal from prior information encoded to capture variations over a population in form, shape, anatomic layout, and image appearance of objects. Model/atlas-based methods are extant in medical image segmentation. Although multi-atlas/ multi-model methods have shown improved accuracy for image segmentation, if the atlases/models do not cover representatively the distinct groups, then the methods may not be generalizable to new populations. In a previous study, we have given an answer to address the following problem at image level: How many models/ atlases are needed for optimally encoding prior information to address the differing body habitus factor in a population? However, the number of models for different objects may be different, and at the image level, it may not be possible to infer the number of models needed for each object. So, the modified question to which we are now seeking an answer to in this paper is: How many models/ atlases are needed for optimally encoding prior information to address the differing body habitus factor for each object in a body region? To answer this question, we modified our method in the previous study for seeking the optimum grouping for a given population of images but focusing on the individual objects. We present our results on head and neck computed tomography (CT) scans of 298 patients.

LinSEM: Linearizing segmentation evaluation metrics for medical images

Numerous algorithms are available for segmenting medical images. Empirical discrepancy metrics are commonly used in measuring the similarity or difference between segmentations by algorithms and "true" segmentations. However, one issue with the commonly used metrics is that the same metric value often represents different levels of "clinical acceptability" for different objects depending on their size, shape, and complexity of form. An ideal segmentation evaluation metric should be able to reflect degrees of acceptability directly from metric values and be able to show the same acceptability meaning by the same metric value for objects of different shape, size, and form. Intuitively, metrics which have a linear relationship with degree of acceptability will satisfy these conditions of the ideal metric. This issue has not been addressed in the medical image segmentation literature. In this paper, we propose a method called LinSEM for linearizing commonly used segmentation evaluation metrics based on corresponding degrees of acceptability evaluated by an expert in a reader study. LinSEM consists of two main parts: (a) estimating the relationship between metric values and degrees of acceptability separately for each considered metric and object, and (b) linearizing any given metric value corresponding to a given segmentation of an object based on the estimated relationship. Since algorithmic segmentations do not usually cover the full range of variability of acceptability, we create a set (SS) of simulated segmentations for each object that guarantee such coverage by using image transformations applied to a set (ST) of true segmentations of the object. We then conduct a reader study wherein the reader assigns an acceptability score (AS) for each sample in SS, expressing the acceptability of the sample on a 1 to 5 scale. Then the metric-AS relationship is constructed for the object by using an estimation method. With the idea that the ideal metric should be linear with respect to acceptability, we can then linearize the metric value of any segmentation sample of the object from a set (SA) of actual segmentations to its linearized value by using the constructed metric-acceptability relationship curve. Experiments are conducted involving three metrics - Dice coefficient (DC), Jaccard index (JI), and Hausdorff Distance (HD) - on five objects: skin outer boundary of the head and neck (cervico-thoracic) body region superior to the shoulders, right parotid gland, mandible, cervical esophagus, and heart. Actual segmentations (SA) of these objects are generated via our Automatic Anatomy Recognition (AAR) method. Our results indicate that, generally, JI has a more linear relationship with acceptability before linearization than other metrics. LinSEM achieves significantly improved uniformity of meaning post-linearization across all tested objects and metrics, except in a few cases where the departure from linearity was insignificant. This improvement is generally the largest for DC and HD reaching 8-25% for many tested cases. Although some objects (such as right parotid gland and esophagus for DC and JI) are close in their meaning between themselves before linearization, they are distant in this meaning from other objects but are brought close to other objects after linearization. This suggests the importance of performing linearization considering all objects in a body region and body-wide.

How many models/atlases are needed as priors for capturing anatomic population variations?

Many medical image processing and analysis operations can benefit a great deal from prior information encoded in the form of models/atlases to capture variations over a population in form, shape, anatomic layout, and image appearance of objects. However, two fundamental questions have not been addressed in the literature: "How many models/atlases are needed for optimally encoding prior information to address the differing body habitus factor in that population?" and "Images of how many subjects in the given population are needed to optimally harness prior information?" We propose a method to seek answers to these questions. We assume that there is a well-defined body region of interest and a subject population under consideration, and that we are given a set of representative images of the body region for the population. After images are trimmed to the exact body region, a hierarchical agglomerative clustering algorithm partitions the set of images into a specified number of groups by using pairwise image (dis)similarity as a cost function. Optionally the images may be pre-registered among themselves prior to this partitioning operation. We define a measure called Residual Dissimilarity (RD) to determine the goodness of each partition. We then ascertain how RD varies as a function of the number of elements in the partition for finding the optimum number(s) of groups. Breakpoints in this function are taken as the recommended number of groups/models/atlases. Our results from analysis of sizeable CT data sets of adult patients from two body regions - thorax (346) and head and neck (298) - can be summarized as follows. (1) A minimum of 5 to 8 groups (or models/atlases) seems essential to properly capture information about differing anatomic forms and body habitus. (2) A minimum of 150 images from different subjects in a population seems essential to cover the anatomical variations for a given body region. (3) In grouping, body habitus variations seem to override differences due to other factors such as gender, with/without contrast enhancement in image acquisition, and presence of moderate pathology. This method may be helpful for constructing high quality models/atlases from a sufficiently large population of images and in optimally selecting the training image sets needed in deep learning strategies.