Head Segmentation Research Articles

CAU-Net: An Improved Attention U-Net for CT Head and Neck Organs at Risk Contouring

Segmentation of head and neck (H&N) organs at risk (OARs) is an intricate process, we propose Contest- extractor Attention U-Net (CAU-Net) to improve the segmentation accuracy of OARs contours by addressing the problem of limited accuracy of 2D segmentation algorithms in current radiotherapy techniques. A total of 60 patients from CSTRO2019's H&N dataset containing 22 organs, were available to train and evaluate a prototype deep learning-based normal tissue 2D auto-segmentation algorithm. Our CAU-Net is based on U-Net, by using the edge attention module to enhance the boundary representation, the null convolution block in the context extraction module to encode high-level semantic feature information, and the convolution of the sensory field to assign to different targets. A Dice loss function combine contour loss function was used in training the models. The contour loss function was improved to segment the target regions by the weights of different organ occurrences and the region assignment of false positives and false negatives to accurately predict the boundary structure. The OARs were delineated by a single experienced physician. A subset of 10 cases was withheld from training and used for validation. On those, we set three different deep-learning networks trained with CSTRO and compared them to the gold data: A) CAU-Net, B) nnUNet, and C) UNet++. To test its applicability, we used another public H&N dataset Public Domain Database for Computational Anatomy (PDDCA) containing 8 organs with 47 patients, among which 10 cases were used for validation: D) CAU-Net with the PDDCA, E) UNet2022 with the PDDCA. The Dice similarity coefficient (DSC) was used to measure the overlap between the results of the gold data and the automated segmentations. The average DSC scores for method A, B, and C across all OARs in the 10 evaluation cases were 0.67±0.08, 0.58±0.11 and 0.62±0.12, respectively. The difference in mean DSC scores was significant (p<0.05). The A/B difference was significant in Lens-L, Lens-R and Pituitary. Method A scored the highest DSC in all OARs except for the Spinal Cord, Mandible-L and Mandible-R. 16 OARs showed DSC≥0.6 on CSTRO. Method D, and E achieved 0.84±0.10 and 0.83±0.09 average DSC respectively. All OARs showed DSC≥0.7 on PDDCA. The CAU-Net proposed by us achieved better results than the baseline network for H&N OAR segmentation. This new development will provide the possibility of H&N organ segmentation and rapid diagnosis of radiotherapy. All the networks trained with PDDCA scored higher than CSTRO. Auto segmentation results can differ significantly when the same algorithm is trained on data from different institutions.

Mask Region-Based Convolutional Neural Network segmentation of the humerus and scapula from proton density-weighted axial shoulder magnetic resonance images.

This study aimed to evaluate the effectiveness of Mask Region-Based Convolutional Neural Network (R-CNN) in humerus and scapula segmentation. The study included 665 axial proton density (PD)-weighted magnetic resonance images of 665 consecutive shoulder instability patients (412 males, 253 females; mean age: 27±5.2 years; range, 18 to 42 years) between January 2011 and December 2014. Mask R-CNN was used to automatically segment humerus and scapula regions simultaneously. Segmentation success of Mask R-CNN was compared to the manual segmentation results of an orthopedic surgeon. Statistical evaluation was done with the Dice coefficient and the mean average precision) score. According to the humeral head structure three groups were generated: the healthy humeral head group, the edematous humeral head group, and the Hill-Sachs group (humeral heads with Hill-Sachs lesions). In the test images, 81 humeral heads were healthy, 100 were edematous, and 38 had a Hill-Sachs lesion. According to the Dice metric, the overall success rate of Mask R-CNN configuration was 96.47 and 93.87% for the segmentation of the humeral head and scapula, respectively, and 95.86 and 92.35% for an intersection over union of 0.5 according to the mean average precision. According to the Dice metric, the segmentation success of the humerus and scapula of the healthy group was 94.58 and 97.42%, the segmentation success of the edematous humerus group was 93.56 and 96.53%, and the segmentation success of the Hill-Sachs group was 93.47 to 95.48%. The segmentation success of scapula in the case of discontinuity was 92.86% according to Dice metric. Mask R-CNN-based humerus and scapula segmentation provided promising results compared to manual segmentation of an expert. Mask R-CNN overcomes the problem of discontinuous edges and Rician noise in axial PD-weighted shoulder magnetic resonance imaging.

Multi-compartment head modeling in EEG: Unstructured boundary-fitted tetra meshing with subcortical structures.

This paper introduces an automated approach for generating a finite element (FE) discretization of a multi-compartment human head model for electroencephalographic (EEG) source localization. We aim to provide an adaptable FE mesh generation tool for EEG studies. Our technique relies on recursive solid angle labeling of a surface segmentation coupled with smoothing, refinement, inflation, and optimization procedures to enhance the mesh quality. In this study, we performed numerical meshing experiments with the three-layer Ary sphere and a magnetic resonance imaging (MRI)-based multi-compartment head segmentation which incorporates a comprehensive set of subcortical brain structures. These experiments are motivated, on one hand, by the sensitivity of non-invasive subcortical source localization to modeling errors and, on the other hand, by the present lack of open EEG software pipelines to discretize all these structures. Our approach was found to successfully produce an unstructured and boundary-fitted tetrahedral mesh with a sub-one-millimeter fitting error, providing the desired accuracy for the three-dimensional anatomical details, EEG lead field matrix, and source localization. The mesh generator applied in this study has been implemented in the open MATLAB-based Zeffiro Interface toolbox for forward and inverse processing in EEG and it allows for graphics processing unit acceleration.

Automatic Segmentation of Head and Neck Cancer from PET-MRI Data Using Deep Learning

PurposeHead and neck squamous cell carcinoma (HNSCC) is one of the most common cancer types globally. Due to the complex anatomy of the region, diagnosis and treatment is challenging. Early diagnosis and treatment are important, because advanced and recurrent HNSCC have a poor prognosis. Robust and precise tools are needed to help diagnose HNSCC reliably in its early stages. The aim of this study was to assess the applicability of a convolutional neural network in detecting and auto-delineating HNSCC from PET-MRI data.Methods2D U-net models were trained and tested on PET, MRI, PET-MRI and augmented PET-MRI data from 44 patients diagnosed with HNSCC. The scans were taken 12 weeks after chemoradiation therapy with a curative intention. A proportion of the patients had follow-up scans which were included in this study as well, giving a total of 62 PET-MRI scans. The scans yielded a total of 178 PET-MRI slices with cancer. A corresponding number of negative slices were chosen randomly yielding a total of 356 slices. The data was divided into training, validation and test sets (n = 247, n = 43 and n = 66 respectively). Dice score was used to evaluate the segmentation accuracy. In addition, the classification capabilities of the models were assessed.ResultsWhen true positive segmentations were considered, the mean Dice scores for the test set were 0.79, 0.84 and 0.87 for PET, PET-MRI and augmented PET-MRI, respectively. Classification accuracies were 0.62, 0.71 and 0.65 for PET, PET-MRI and augmented PET-MRI, respectively. The MRI based model did not yield segmentation results. A statistically significant difference was found between the PET-MRI and PET models (p = 0.008).ConclusionAutomatic segmentation of HNSCC from the PET-MRI data with 2D U-nets was shown to give sufficiently accurate segmentations.

Convolutional Neural Network for Head Segmentation and Counting in Crowded Retail Environment Using Top-view Depth Images

Convolutional Neural Network for Head Segmentation and Counting in Crowded Retail Environment Using Top-view Depth Images

MHL-Net: A Multistage Hierarchical Learning Network for Head and Neck Multiorgan Segmentation.

Accurate segmentation of head and neck organs at risk is crucial in radiotherapy. However, the existing methods suffer from incomplete feature mining, insufficient information utilization, and difficulty in simultaneously improving the performance of small and large organ segmentation. In this paper, a multistage hierarchical learning network is designed to fully extract multidimensional features, combined with anatomical prior information and imaging features, using multistage subnetworks to improve the segmentation performance. First, multilevel subnetworks are constructed for primary segmentation, localization, and fine segmentation by dividing organs into two levels-large and small. Different networks both have their own learning focuses and feature reuse and information sharing among each other, which comprehensively improved the segmentation performance of all organs. Second, an anatomical prior probability map and a boundary contour attention mechanism are developed to address the problem of complex anatomical shapes. Prior information and boundary contour features effectively assist in detecting and segmenting special shapes. Finally, a multidimensional combination attention mechanism is proposed to analyze axial, coronal, and sagittal information, capture spatial and channel features, and maximize the use of structural information and semantic features of 3D medical images. Experimental results on several datasets showed that our method was competitive with state-of-the-art methods and improved the segmentation results for multiscale organs.

Automatic Segmentation of the Paediatric Femoral Head.

Developmental dysplasia of the hip (DDH) is a developmental deformity occurring in 0.1-3.4% of infants. Timely surgical intervention can ameliorate the condition in stable hips and reduce future cases of osteoarthritis and total hip replacement. However, current definitions of DDH are subjective, and thus would benefit from a more objective and reliable assessment metric. Since the shape of the femoral head and its congruence with the acetabulum are disrupted by DDH, analysis of the femoral head could potentially play a role in the development of novel objective morphological metric for stable DDH. Therefore, this paper aimed to segment the paediatric femoral head in stable hips from radiographs, which has not been attempted before in the chosen focus age group (1-16 years) where the pelvis and hip joint undergo significant development. Two techniques were compared against a baseline U-Net: data augmentation and region-of-interest (ROI) networks. Four models were developed either without, with just one, or with both techniques. Evaluated using tenfold cross-validation, the U-Net trained with both techniques achieved the best results, with a Dice Similarity Coefficient (DSC) of 0.951±0.037 (mean ± standard deviation, calculated with 720 images). Future work will use this segmentation algorithm to accurately characterise hip joint morphology and estimate the benefit of early surgical intervention in DDH.

Prompt-Based Tuning of Transformer Models for Multi-Center Medical Image Segmentation of Head and Neck Cancer.

Medical image segmentation is a vital healthcare endeavor requiring precise and efficient models for appropriate diagnosis and treatment. Vision transformer (ViT)-based segmentation models have shown great performance in accomplishing this task. However, to build a powerful backbone, the self-attention block of ViT requires large-scale pre-training data. The present method of modifying pre-trained models entails updating all or some of the backbone parameters. This paper proposes a novel fine-tuning strategy for adapting a pretrained transformer-based segmentation model on data from a new medical center. This method introduces a small number of learnable parameters, termed prompts, into the input space (less than 1% of model parameters) while keeping the rest of the model parameters frozen. Extensive studies employing data from new unseen medical centers show that the prompt-based fine-tuning of medical segmentation models provides excellent performance regarding the new-center data with a negligible drop regarding the old centers. Additionally, our strategy delivers great accuracy with minimum re-training on new-center data, significantly decreasing the computational and time costs of fine-tuning pre-trained models. Our source code will be made publicly available.

Information fusion for fully automated segmentation of head and neck tumors from PET and CT images.

PET/CT images combining anatomic and metabolic data provide complementary information that can improve clinical task performance. PET image segmentation algorithms exploiting the multi-modal information available are still lacking. Our study aimed to assess the performance of PET and CT image fusion for gross tumor volume (GTV) segmentations of head and neck cancers (HNCs) utilizing conventional, deep learning (DL), and output-level voting-based fusions. The current study is based on a total of 328 histologically confirmed HNCs from six different centers. The images were automatically cropped to a 200 × 200 head and neck region box, and CT and PET images were normalized for further processing. Eighteen conventional image-level fusions were implemented. In addition, a modified U2-Net architecture as DL fusion model baseline was used. Three different input, layer, and decision-level information fusions were used. Simultaneous truth and performance level estimation (STAPLE) and majority voting to merge different segmentation outputs (from PET and image-level and network-level fusions), that is, output-level information fusion (voting-based fusions) were employed. Different networks were trained in a 2D manner with a batch size of 64. Twenty percent of the dataset with stratification concerning the centers (20% in each center) were used for final result reporting. Different standard segmentation metrics and conventional PET metrics, such as SUV, were calculated. In single modalities, PET had a reasonable performance with a Dice score of 0.77±0.09, while CT did not perform acceptably and reached a Dice score of only 0.38±0.22. Conventional fusion algorithms obtained a Dice score range of [0.76-0.81] with guided-filter-based context enhancement (GFCE) at the low-end, and anisotropic diffusion and Karhunen-Loeve transform fusion (ADF), multi-resolution singular value decomposition (MSVD), and multi-level image decomposition based on latent low-rank representation (MDLatLRR) at the high-end. All DL fusion models achieved Dice scores of 0.80. Output-level voting-based models outperformed all other models, achieving superior results with a Dice score of 0.84 for Majority_ImgFus, Majority_All, and Majority_Fast. A mean error of almost zero was achieved for all fusions using SUVpeak , SUVmean and SUVmedian . PET/CT information fusion adds significant value to segmentation tasks, considerably outperforming PET-only and CT-only methods. In addition, both conventional image-level and DL fusions achieve competitive results. Meanwhile, output-level voting-based fusion using majority voting of several algorithms results in statistically significant improvements in the segmentation of HNC.

Precise segmentation of fetal head in ultrasound images using improved U‐Net model

AbstractMonitoring fetal growth in utero is crucial to anomaly diagnosis. However, current computer‐vision models struggle to accurately assess the key metrics (i.e., head circumference and occipitofrontal and biparietal diameters) from ultrasound images, largely owing to a lack of training data. Mitigation usually entails image augmentation (e.g., flipping, rotating, scaling, and translating). Nevertheless, the accuracy of our task remains insufficient. Hence, we offer a U‐Net fetal head measurement tool that leverages a hybrid Dice and binary cross‐entropy loss to compute the similarity between actual and predicted segmented regions. Ellipse‐fitted two‐dimensional ultrasound images acquired from the HC18 dataset are input, and their lower feature layers are reused for efficiency. During regression, a novel region of interest pooling layer extracts elliptical feature maps, and during segmentation, feature pyramids fuse field‐layer data with a new scale attention method to reduce noise. Performance is measured by Dice similarity, mean pixel accuracy, and mean intersection‐over‐union, giving 97.90%, 99.18%, and 97.81% scores, respectively, which match or outperform the best U‐Net models.

A regressive encoder-decoder-based deep attention model for segmentation of fetal head in 2D-ultrasound images

Ultrasound imaging is the most commonly used imaging during pregnancy for tracking the fetus's growth and monitoring other biological parameters. The assessment of the development of the baby's growth requires imaging-based analysis in every trimester. The automatic computerized software and systems provide the platform for radiologists to more accurately access the fetus's head circumference as compared to manual estimation. The improvement of such computerized algorithms is always the key demand to improve accuracy and precision. This paper proposes an improved encoder-decoder model for the segmentation of the fetal head segmentation in 2D-ultrasound images. The proposed model uses regression in combination with attention to the encoder-decoder model to determine the fetus's head circumference. The model is further extended with the post-processing ellipse fitting to superimpose the segmentation region on ultrasound images for clear visualization of the fetus's head. Further, the proposed model performance is evaluated by using various statistical measures using segmented regions and available ground truth. The experimental results demonstrate a similarity score of 94.56%. The comparative result suggests that the proposed model is providing a more accurate fetus head segmentation region on 2D-ultrasound images as compared to other existing approaches.

DMCT-Net: dual modules convolution transformer network for head and neck tumor segmentation in PET/CT

Objective. Accurate segmentation of head and neck (H&N) tumors is critical in radiotherapy. However, the existing methods lack effective strategies to integrate local and global information, strong semantic information and context information, and spatial and channel features, which are effective clues to improve the accuracy of tumor segmentation. In this paper, we propose a novel method called dual modules convolution transformer network (DMCT-Net) for H&N tumor segmentation in the fluorodeoxyglucose positron emission tomography/computed tomography (FDG-PET/CT) images. Approach. The DMCT-Net consists of the convolution transformer block (CTB), the squeeze and excitation (SE) pool module, and the multi-attention fusion (MAF) module. First, the CTB is designed to capture the remote dependency and local multi-scale receptive field information by using the standard convolution, the dilated convolution, and the transformer operation. Second, to extract feature information from different angles, we construct the SE pool module, which not only extracts strong semantic features and context features simultaneously but also uses the SE normalization to adaptively fuse features and adjust feature distribution. Third, the MAF module is proposed to combine the global context information, channel information, and voxel-wise local spatial information. Besides, we adopt the up-sampling auxiliary paths to supplement the multi-scale information. Main results. The experimental results show that the method has better or more competitive segmentation performance than several advanced methods on three datasets. The best segmentation metric scores are as follows: DSC of 0.781, HD95 of 3.044, precision of 0.798, and sensitivity of 0.857. Comparative experiments based on bimodal and single modal indicate that bimodal input provides more sufficient and effective information for improving tumor segmentation performance. Ablation experiments verify the effectiveness and significance of each module. Significance. We propose a new network for 3D H&N tumor segmentation in FDG-PET/CT images, which achieves high accuracy.

Modularity, homology, heterochrony: Gavin de Beer's legacy to the mammalian skull.

Modularity (segmentation), homology and heterochrony were essential concepts embraced by Gavin de Beer in his studies of the development and evolution of the vertebrate skull. While his pioneering contributions have stood the test of time, our understanding of the biological processes that underlie each concept has evolved. We assess de Beer's initial training as an experimental embryologist; his switch to comparative and descriptive studies of skulls, jaws and middle ear ossicles; and his later research on the mammalian skull, including his approach to head segmentation. The role of cells of neural crest and mesodermal origin in skull development, and developmental, palaeontological and molecular evidence for the origin of middle ear ossicles in the evolutionary transition from reptiles to mammals are used to illustrate our current understanding of modularity, homology and heterochrony. This article is part of the theme issue 'The mammalian skull: development, structure and function'.

MultiTrans: Multi-scale feature fusion transformer with transfer learning strategy for multiple organs segmentation of head and neck CT images

Radiotherapy with precise segmentation of head and neck organs at risk (OARs) is one of the important treatment methods for head and neck cancer. In routine clinical practice, OARs are manually segmented by doctors to avoid irreversible adverse reactions caused by radiotherapy, which is time-consuming and laborious. To assist doctors in OARs segmentation, a MultiTrans framework with a multi-scale feature fusion module was proposed in this paper. In the multi-scale feature fusion module, the original image and the feature map of CNN were fused together to form a compound feature map for more complete high-resolution global information. In addition, the global information was also fully utilized in MultiTrans by using the feature map restored from the compound feature map in the skip connection. The multi-scale interactive high-resolution information can make full use of medical image information and obtain features more comprehensively, thus improve the segmentation accuracy. Experiments showed that MultiTrans had an average Dice score coefficient (DSC) of 74.01% in all organs, effectively improved segmentation accuracy. In addition, we proposed a transfer learning strategy for small organs by transferring the weight parameters of organs with a large amount of data to organs with a small amount of data to speed up the convergence of MultiTrans and reduce the demand for data volume in the MultiTrans. With this strategy, the average DSC of small organs was obviously increased, making the segmentation of small organs more accurate. The proposed framework and transfer learning strategy have the potential of assisting doctors in OARs delineation.

MO-0799 Single-click user input reduces false detection in deep learning head and neck tumor segmentation

MO-0799 Single-click user input reduces false detection in deep learning head and neck tumor segmentation

Hybrid neural-like P systems with evolutionary channels for multiple brain metastases segmentation

Neural-like P systems are membrane computing models inspired by natural computing. Spiking neurral (SN) P systems, a kind of neural-like P systems, are viewed as third-generation neural network models. Although real neurons have complex structures, classical SN P systems simplify the structures and corresponding mechanisms to stationary two-dimensional graphs and lack related evolution mechanisms on spikes and channels, which limits the real applications of these models. In this paper, we propose a new hybrid SN P system with evolutionary channels (HN P systems), including three new types of rules for dynamically generating or removing one-one and one-many/many-one channels with related evolutions of spikes on the hybrid neuron structures. Two dynamic regulatory factors are also presented on rules to help guide the optimization of the HN P systems automatically. Based on the new P system, a multiple brain metastases (BMs) segmentation model is developed. The experimental results indicate that the proposed models outperform the state-of-the-art methods on the BMs, which have large variations in sizes, positions and shapes, and low contrast with their surroundings. Performances on the head and neck segmentation dataset also verifies the effectiveness of the HN P system.

Head and neck tumor segmentation convolutional neural network robust to missing PET/CT modalities using channel dropout

Objective. Radiation therapy for head and neck (H&N) cancer relies on accurate segmentation of the primary tumor. A robust, accurate, and automated gross tumor volume segmentation method is warranted for H&N cancer therapeutic management. The purpose of this study is to develop a novel deep learning segmentation model for H&N cancer based on independent and combined CT and FDG-PET modalities. Approach. In this study, we developed a robust deep learning-based model leveraging information from both CT and PET. We implemented a 3D U-Net architecture with 5 levels of encoding and decoding, computing model loss through deep supervision. We used a channel dropout technique to emulate different combinations of input modalities. This technique prevents potential performance issues when only one modality is available, increasing model robustness. We implemented ensemble modeling by combining two types of convolutions with differing receptive fields, conventional and dilated, to improve capture of both fine details and global information. Main Results. Our proposed methods yielded promising results, with a Dice similarity coefficient (DSC) of 0.802 when deployed on combined CT and PET, DSC of 0.610 when deployed on CT, and DSC of 0.750 when deployed on PET. Significance. Application of a channel dropout method allowed for a single model to achieve high performance when deployed on either single modality images (CT or PET) or combined modality images (CT and PET). The presented segmentation techniques are clinically relevant to applications where images from a certain modality might not always be available.

Realistic CT data augmentation for accurate deep-learning based segmentation of head and neck tumors in kV images acquired during radiation therapy.

Using radiation therapy (RT) to treat head and neck (H&N) cancers requires precise targeting of the tumor to avoid damaging the surrounding healthy organs. Immobilisation masks and planning target volume margins are used to attempt to mitigate patient motion during treatment, however patient motion can still occur. Patient motion during RT can lead to decreased treatment effectiveness and a higher chance of treatment related side effects. Tracking tumor motion would enable motion compensation during RT, leading to more accurate dose delivery. The purpose of this paper is to develop a method to detect and segment the tumor in kV images acquired during RT. Unlike previous tumor segmentation methods for kV images, in this paper a process for generating realistic and synthetic CT deformations was developed to augment the training data, and make the segmentation method robust to patient motion. Detecting the tumor in 2D kV images is a necessary step towards 3D tracking of the tumor position during treatment. In this paper a conditional generative adversarial network (cGAN) is presented that can detect and segment the gross tumor volume (GTV) in kV images acquired during H&N RT. Retrospective data from 15 H&N cancer patients obtained from the Cancer Imaging Archive were used to train and test patient-specific cGANs. The training data consisted of digitally reconstructed radiographs (DRRs) generated from each patient's planning CT and contoured GTV. Training data was augmented by using synthetically deformed CTs to generate additional DRRs (in total 39600 DRRs per patient or 25200 DRRs for nasopharyngeal patients) containing realistic patient motion. The method for deforming the CTs was a novel deformation method based on simulating head rotation and internal tumor motion. The testing dataset consisted of 1080 DRRs for each patient, obtained by deforming the planning CT and GTV at different magnitudes to the training data. The accuracy of the generated segmentations was evaluated by measuring the segmentation centroid error, Dice similarity coefficient (DSC) and mean surface distance (MSD). This paper evaluated the hypothesis that when patient motion occurs, using a cGAN to segment the GTV would create a more accurate segmentation than no-tracking segmentations from the original contoured GTV, the current standard-of-care. This hypothesis was tested using the 1-tailed Mann-Whitney U-test. The magnitude of our cGAN segmentation centroid error was (mean ± standard deviation) 1.1 ± 0.8 millimeters and the DSC and MSD values were 0.90 ± 0.03 and 1.6 ± 0.5 millimeters respectively. Our cGAN segmentation method reduced the segmentation centroid error (p<0.001), and MSD (p = 0.031) when compared to the no-tracking segmentation, but did not significantly increase the DSC (p = 0.294). The accuracy of our cGAN segmentation method demonstrates the feasibility of this method for H&N cancer patients during RT. Accurate tumor segmentation of H&N tumors would allow for intrafraction monitoring methods to compensate for tumor motion during treatment, ensuring more accurate dose delivery and enabling better H&N cancer patient outcomes. This article is protected by copyright. All rights reserved.

Accurate segmentation of head and neck radiotherapy CT scans with 3D CNNs: consistency is key

Objective. Automatic segmentation of organs-at-risk in radiotherapy planning computed tomography (CT) scans using convolutional neural networks (CNNs) is an active research area. Very large datasets are usually required to train such CNN models. In radiotherapy, large, high-quality datasets are scarce and combining data from several sources can reduce the consistency of training segmentations. It is therefore important to understand the impact of training data quality on the performance of auto-segmentation models for radiotherapy. Approach. In this study, we took an existing 3D CNN architecture for head and neck CT auto-segmentation and compare the performance of models trained with a small, well-curated dataset (n = 34) and then a far larger dataset (n = 185) containing less consistent training segmentations. We performed 5-fold cross-validations in each dataset and tested segmentation performance using the 95th percentile Hausdorff distance and mean distance-to-agreement metrics. Finally, we validated the generalisability of our models with an external cohort of patient data (n = 12) with five expert annotators. Main results. The models trained with a large dataset were greatly outperformed by models (of identical architecture) trained with a smaller, but higher consistency set of training samples. Our models trained with a small dataset produce segmentations of similar accuracy as expert human observers and generalised well to new data, performing within inter-observer variation. Significance. We empirically demonstrate the importance of highly consistent training samples when training a 3D auto-segmentation model for use in radiotherapy. Crucially, it is the consistency of the training segmentations which had a greater impact on model performance rather than the size of the dataset used.

Automatic gross tumor segmentation of canine head and neck cancer using deep learning and cross-species transfer learning.

Radiotherapy (RT) is increasingly being used on dogs with spontaneous head and neck cancer (HNC), which account for a large percentage of veterinary patients treated with RT. Accurate definition of the gross tumor volume (GTV) is a vital part of RT planning, ensuring adequate dose coverage of the tumor while limiting the radiation dose to surrounding tissues. Currently the GTV is contoured manually in medical images, which is a time-consuming and challenging task. The purpose of this study was to evaluate the applicability of deep learning-based automatic segmentation of the GTV in canine patients with HNC. Contrast-enhanced computed tomography (CT) images and corresponding manual GTV contours of 36 canine HNC patients and 197 human HNC patients were included. A 3D U-Net convolutional neural network (CNN) was trained to automatically segment the GTV in canine patients using two main approaches: (i) training models from scratch based solely on canine CT images, and (ii) using cross-species transfer learning where models were pretrained on CT images of human patients and then fine-tuned on CT images of canine patients. For the canine patients, automatic segmentations were assessed using the Dice similarity coefficient (Dice), the positive predictive value, the true positive rate, and surface distance metrics, calculated from a four-fold cross-validation strategy where each fold was used as a validation set and test set once in independent model runs. CNN models trained from scratch on canine data or by using transfer learning obtained mean test set Dice scores of 0.55 and 0.52, respectively, indicating acceptable auto-segmentations, similar to the mean Dice performances reported for CT-based automatic segmentation in human HNC studies. Automatic segmentation of nasal cavity tumors appeared particularly promising, resulting in mean test set Dice scores of 0.69 for both approaches. In conclusion, deep learning-based automatic segmentation of the GTV using CNN models based on canine data only or a cross-species transfer learning approach shows promise for future application in RT of canine HNC patients.