Abstract
Background:
Deep learning has significantly improved medical image segmentation, particularly in complex applications such as gynecological brachytherapy (BT). This study proposes a novel U-Net variant, the residual in residual dense block U-Net (RU-Net), to address the segmentation of organs-at-risk (OARs), which are often distorted due to BT probe insertion. Accurate segmentation is essential for effective treatment planning; however, task-specific datasets are scarce. To tackle this, a dedicated magnetic resonance imaging (MRI) dataset was developed, focusing on the precise delineation of OARs.
Methods:
The RU-Net architecture enhances feature extraction through embedded residual in residual dence block (RRDB) modules. It was trained and evaluated using a multiplanar (axial, coronal, and sagittal) MRI dataset. The study benchmarks RU-Net against pyramid scene parsing network, residual U-Net, U-Net++, and traditional U-Net models. Performance was assessed using categorical cross-entropy and dice loss metrics.
Results:
Under the dice loss function, RU-Net consistently outperformed competing models across all anatomical planes. It achieved superior scores in accuracy, precision, recall, and intersection over union (IOU), particularly excelling in class-wise IOUs for critical structures such as bladder, rectum, sigmoid colon, and femoral heads.
Discussion:
The architecture improved both boundary delineation and overall prediction quality. It is established that the proposed RU-Net shows its supremacy under dice loss conditions in terms of proper prediction and segmentation.
Conclusion:
RU-Net offers a robust solution for multiorgan segmentation in gynecological BT. Its advanced design effectively addresses anatomical distortions, demonstrating strong potential to improve clinical outcomes in magnetic resonance-guided treatment planning.
Keywords: Deep-learning, magnetic resonance-guided brachytherapy, multiclass segmentation, residuals in residual dense block
INTRODUCTION
Background
Cervical cancer ranks as the fourth-most common cancer affecting women worldwide, with an estimated 342,000 deaths in 2020. The highest incidence and mortality rates are found in low socioeconomic or developing countries due to limited access to screening and various social and economic factors.[1,2] For the radical treatment of locally advanced tumors, a combination of external beam radiotherapy (EBRT) and brachytherapy (BT) is employed. In BT, a radioisotope is placed inside the target area to achieve local tumor control. BT has been demonstrated to be indispensable and is strongly associated with a greater survival rate. Some studies indicate that BT can improve the survival rate of cervical cancer by 40% compared to using EBRT alone.[3]
Because of the significant dose gradient in BT, contouring of organs at risk (OARs) should be done cautiously and appropriately in HDR-BT treatment to improve organ sparing and tumor control. However, unlike in EBRT, there is limited time for planning in HDR-BT because radiation oncologists and medical physicists must complete contouring as quickly as possible to reduce patient discomfort and movement. Manually delineating these images is a tedious and exhausting process that can delay treatment outcomes and does not adequately meet clinical needs.[4] In addition, manual segmentation results can vary among radiation oncologists, leading to inconsistencies in the anatomical borders of critical organs at risk. These interobserver variations can cause dosimetry discrepancies, potentially leading to complications and affecting the outcomes in BT treatment planning.[5]
Computerized medical image segmentation of human organs from magnetic resonance imaging (MRI) data has the potential to assist radiation oncologists in their daily tasks related to BT planning.[6] Computationally achieved, precise, rapid, and automated segmentation methods for multiorgan segmentation, offering reduced waiting times, consistency, and high reproducibility, have demonstrated clinical feasibility and utmost importance for patients.[7] Conventional clinical image segmentation, which serves as the benchmark in medical image analysis, typically involves custom-built image feature extraction, such as anatomical border detection and mathematical models to track image gradients along object boundaries. Autosegmentation tools for OARs can enhance radiation oncologists’ work and minimize variability.[8,9,10,11] In head and neck cancer, atlas-based autosegmentation (ABAS) algorithms are commonly used for identifying organs at risk.[12] However, in female gynecological malignancies, the critical organs often exhibit significant irregularities in shape, size, and location and are poorly distinguishable from tumors.[13] ABAS techniques, which segment contours based on a library of reference images and map elements to the target image using a deformable image registration algorithm, have been popular in the last 10 years. Kim et al.[14] segmented OARs in endometrial gynecological cancer, achieving a best dice of 0.75 and an average segmentation duration of 45.1 s. Although ABAS improves contouring efficiency, it also has significant drawbacks. Kim et al.[14] claimed that organs that are isodense with their environment are unsuitable for ABAS. Teguh et al.[15] found that ABAS performs poorly for small and thin OARs. Generally, computed tomography (CT) is the preferred imaging method for BT target and organ delineation, as well as applicator localization.[16] However, MRI is highly recommended for delineating tumors and adjacent critical organs, as it provides superior soft-tissue contrast compared to CT, thereby improving accuracy and treatment outcomes.[17] Once BT planning images are sent to the treatment planning system, the initial step is to identify the target volume and its adjacent critical organs, a process typically done manually by the attending physician, making it very labor-intensive and time-consuming.
Despite advancements in medical image analysis, the development of robust multisegmentation algorithms for MRI pelvis images in the context of BT remains an active area of research.[18] Autosegmentation algorithms addressing this area are crucial for reducing variability and processing time.[12] Challenges arise from the complex anatomical variations and imaging artifacts present in pelvic MRI scans, including partial volume effects, organ deformation, and image distortion,[19] severely affect the performance of autosegmentation algorithms in practice.
Literature review
MRI image segmentation using deep learning techniques has emerged as a transformative approach in BT, offering unparalleled precision in target delineation and treatment planning. By harnessing the power of artificial intelligence, deep learning algorithms have revolutionized segmentation, enabling personalized treatment strategies and elevating patient care standards. Previous studies have underscored the pivotal role of deep learning-based MRI segmentation in BT optimization. For instance, the work of Zhang et al.[20] demonstrated the efficacy of convolutional neural networks in automating the segmentation of prostate and surrounding structures on MRI images. Through comprehensive validation studies, the authors showcased the superior accuracy and efficiency of deep learning models compared to conventional segmentation methods, thereby enabling more precise dose optimization and sparing of critical structures in prostate BT.
Furthermore, the study by Li et al.[21] showcased the potential of deep learning algorithms in facilitating adaptive planning strategies in gynecological BT. By leveraging recurrent neural networks for real-time segmentation of tumor volumes and OARs on intraoperative MRI scans, the authors demonstrated the feasibility of adaptive treatment modifications based on dynamic anatomical changes. Such advancements not only enhance treatment precision but also minimize the risk of underdosing or overexposure to radiation, thereby improving therapeutic outcomes and patient safety. In addition to treatment planning, deep learning-based MRI segmentation plays a crucial role in longitudinal monitoring of treatment and response assessment. The study by Wang et al.[22] demonstrated the utility of deep learning models for tracking tumor regression and recurrence after BT for cervical cancer. By automating the segmentation of serial MRI scans, the authors enabled timely detection of treatment response dynamics, facilitating prompt adjustments to treatment strategies and optimizing patient outcomes. Comelli et al.[23] present a deep learning-based approach for automated MRI segmentation in prostate cancer imaging, showcasing the potential of artificial intelligence in improving segmentation accuracy and workflow efficiency. Li et al.[24] present the potential of deep learning algorithms in facilitating adaptive planning strategies in gynecological BT. Their study highlights enhanced treatment precision and patient safety through real-time segmentation on intraoperative MRI scans. Hu et al.[25] investigate the feasibility of deep learning-based MRI segmentation in ovarian cancer BT, demonstrating promising results in tumor delineation and treatment planning optimization. Ruiyan Ni et al. studied on the generalizability of deep learning for organ-at-risk segmentation in cervical BT demonstrates the effectiveness of transfer learning to improve model performance across diverse patient datasets.[26] Samuel et al. showed that combining multiple imaging modalities with advanced augmentation techniques significantly enhances segmentation accuracy and model robustness. Studies emphasize the superior performance of integrated approaches over single modality or basic augmentation methods.[27] Based on the literature review, the following research gap needs to be addressed.
Research gap and novelty
Though several methods have been proposed for organ segmentation on MR images, to the best of our knowledge, there is little work that can perform the task in the presence of a foreign object in the body. However, several critical medical diagnoses and treatments demand the insertion of foreign objects inside the body that not only change organ shapes but also change the scanned images depending on the shapes and electromagnetic properties of the foreign objects. In gynecological BT, three applicators or multiple needles with radioactive isotopes are temporarily placed inside the body, which causes a risk to the other vital organs present in the lower abdomen. Evidently, it is difficult to design a model trained on such data that can assist the treatment process by segmenting the vital organs in real time to help avoid unnecessary radiation exposure.
Contributions
This study addresses these challenges by proposing a novel multisegmentation approach tailored specifically for MRI pelvis images in BT planning. Leveraging state-of-the-art deep learning techniques and advanced image processing algorithms, our method aims to accurately delineate relevant anatomical structures while ensuring efficient integration into existing clinical workflows. The major contributions of this research work are as follows:
We propose a novel architecture based on a U-Net backbone, with several residual dense blocks (RDBs), to achieve a fully automatic framework for simultaneous multiorgan segmentation
An investigation on the choice of appropriate loss functions shows that when paired with advanced models such as the RRDB, we can improve the accuracy and efficiency of segmentation tasks, resulting in better treatment planning and outcomes for gynecological BT patients.
Using a recent dataset[28] of pelvic MRI scans, we aim to develop a robust segmentation framework that not only improves accuracy but also enhances efficiency and scalability in clinical practice. The primary objective of this research is to demonstrate the feasibility and efficacy of our proposed multisegmentation framework in improving the quality and consistency of OAR delineation for BT treatment planning. Through comprehensive delineation, evaluation, and validation on a recent dataset of pelvic MRI scans, we seek to establish the clinical utility and potential impact of our approach on patient outcomes, treatment efficacy, and resource optimization in BT practice.
In contrast to traditional segmentation studies, this work specifically addresses the challenging clinical scenario in which foreign objects, such as Fletcher–Suit applicators and intracavitary probes, distort pelvic anatomy and alter MRI characteristics, making organ delineation particularly difficult. The novelty of this work lies in the development of a highly specialized deep learning framework, namely the residual in residual dense block U-Net (RU-Net), for accurate multiorgan segmentation in MR-guided gynecological BT. In contrast to traditional segmentation investigations, our work focuses on the problematic clinical situation where foreign devices such as intracavitary probes and Fletcher–Suit applicators deform the pelvic architecture and change MRI properties, making organ delineation extremely challenging.
In this paper, we present the methodology, experimental setup, results, and discussion of our multisegmentation approach, highlighting its strengths, limitations, and future directions for further refinement and translation into clinical practice. The rest of the paper is organized as follows: The discussion on the model creation and methods is presented in Section 2. The detailed results are presented in Section 3. Following the previous, extensive result analysis is portrayed in Section 4. Finally, in Section 5, the concluding remarks are presented.
METHODS
This section outlines the methodology for developing the proposed RRDB-based U-Net (RU-Net) architecture, experimental setup, evaluation metrics, and loss functions.
Segmentation method
Proposed residuals in residual dense block-based U-Net architecture
Although U-Net models are widely popular in the segmentation tasks, in the U-Net backbone, the decoder receives the features directly from the encoder of the same resolution. This approach might not produce sufficiently good results in complex tasks.[29] We identified that for multiorgan segmentation tasks, the feature at the lowest dimension plays the most crucial role. Thus, in this work, an attempt is made to introduce a modified network from the conventional U-Net network, which consists of two main segments, i.e., a contracting path consisting of the encoders el and an expansive path consisting of the decoders dl; l ∈ {1,2,... L}.[30] The encoders extract higher-order features, followed by a downsampling step, and the encoder combines them and maps them to a higher resolution. Skip connections are utilized in between el and dk of the same spatial dimension. Furthermore, the two portions are connected with the help of a bridge network, which consists of a convolution network. However, in the proposed architecture, the main difference from the conventional U-Net network lies in the bridge section, which connects the contracting and expansive paths. In our approach, the bridge between the two portions is replaced with an RRDB network, depicted in Figure 1. Mathematically, unlike traditional U-Net-based segmentation networks, that are 2 L + 1 layer deep (L layers in contracting path, L layers in decoding path, and a final output layer), our model is 2 L + rbc + 1 layer deep, where the additional rbc layers are introduced by r number of RRDB modules. Each RRDB module contains b dense blocks, and each dense block contains c convolutional layers. However, unlike the recent modified U-Net-based models that introduce the convolutional layer at each feature resolution,[29,31] we proposed to place the convolutional layers at the lowest feature resolution, or at the bridge between the final encoder block eL and the first decoder block d1. However, simply placing convolutional layers causes the vanishing gradient problem and leads to inferior performance. Thus, we design RRDB using residual connections and convolutional layers to mitigate vanishing gradients.
Figure 1.
Proposed RU-Net architecture
In a dense block, output of each intermediate layer θj of dense block θ is defined as:
θj = ϕ(ψ(f1(θj − 1+ θj − 2+... + θ0)) (1)
where θ0 is the input to the dense block, ϕ is the nonlinearity, ψ is the batch normalization (BN) layer, and f1 is a convolutional layer with stride 1. We have considered ϕ as a leaky-ReLU activation in our work. If there are c convolutional layers present in each dense block, the final output of the dense block θc = f1(θc − 1+ θc − 2+... + θ0). For a single RRDB module, the output of each dense block is scaled by 0< β <1 to restrict gradient flow and prevent exploding gradients, as a large number of gradients accumulate in each dense block during backpropagation. If ξk is the input of each intermediate dense block, then
ξk = ξk – 1+ β.θk − 1 (2)
where θk represents the output of the kth dense block. If xi is the input to the ith RRDB module consisting of b dense blocks, the output of that RRDB module is defined as
ωi = xi + β.θbi (3)
where θbi represents the output of the final dense block of the ith RRDB module. If there are r number of RRDB modules present on the bridge, the information flow from eL to d1 is
γe → d = f1 (x1+ ωr) (4)
where x1 is the input to the RRDB module, which is nothing but the output of eL.
In this work, 12 numbers of RRDB networks are considered. Based on Figure 1, the structure of the RRDB can be visualized efficiently. Each RRDB network consists of three dense blocks, where each dense block consists of 4 numbers of 3 × 3 convolution network with stride of 1, followed by a BN layer and Leaky ReLU activation. The concatenation is utilized to combine the signals from the input to the output of the 4th convolution network. The last convolution network consists of a 3 × 3 convolution network with stride 1 and the output signal is scaled by the β = 0.2 parameter. We have used a 2 × 2 maxpooling layer with stride 2 for downsampling in the contracting path, and transposed convolution in the decoder to upsample the images.
An expansive path is used to upsample the images. In this regard, a 2 × 2 transpose convolution layer is taken with stride 2. According to the proposed architecture, the upsample data are concatenated with the corresponding downsample data in the contracting path. This is done to extract more accurate image features. The output layer of the proposed network consists of the number of classes with stride 1 network with pixel-wise soft-max activation.
Experimental setup
To demonstrate the performance of the pyramid scene parsing network (PSP-Net),[32] residual U-Net (ResU-Net),[33] U-Net++,[29] U-Net,[30] and proposed RU-Net models, we have tested them in our dataset with 95 patients. These include bladder, rectum, sigmoid colon, and femoral head from 2D images in various planes (axial, coronal, and sagittal). For this implementation, the Python 3.10 framework is used on a single graphics processing unit machine with 32GB of RAM and NVIDIA Quadro P600 with 16GB of memory.
Evaluation metrics
Performance metrics are essential to evaluate the performance analysis of any machine learning algorithm. Hence, different performance metrics are Accuracy, Recall, Specificity, Precision, and F1-Score.[34] The mathematical expressions of the performance metrics are presented with the help of (5) to (9). In these equations, the terminologies can be explained as true positive (TP), true negative (TN), false positive (FP), and false negative (FN).
Loss functions
Loss functions play an important role in the performance of machine learning algorithms. Based on the objective of the work, the loss functions are decided. In this work, multi-class segmentation issues are addressed. Therefore, this work incorporates two loss functions, namely categorical cross-entropy (CCE)[34,35] and dice loss,[36,37] to conduct a detailed performance analysis. The detailed investigation is carried out using five deep learning models under the abovementioned losses. The mathematical expressions of the CCE and dice loss functions are mentioned in (10) and (11), respectively. In this regard, the detailed analysis and performance assessment are presented in Section 3.
In the above equation, the ground truth and the prediction are represented as κ and κˆ, respectively. The individual pixels are denoted by i, and the classes are denoted by j. In this case, n2 represents the n × n pixels of the image and “C” represents the total number of classes.
RESULTS
In this work, the result analysis is conducted in threefold to establish the proper justifications for utilizing the proposed architecture. The detailed investigations consider the 2D image slices of the axial, coronal, and sagittal planes separately, from a recent clinical MR image dataset comprising 95 patients who underwent intracavitary BT.[38] We delineate the OAR in these scans using ITK-SNAP and prepare it for the MR image segmentation task. For the comparative analysis, five deep-learning models are considered, e.g. PSP-Net, ResU-Net, U-Net++, U-Net, and RU-Net (proposed).[39] To perform this work, the “He-Initializer” is utilized with “Adam” optimization, a 32-batch size with 150 epochs for the investigation in every plane. The CCE and dice loss are considered to be the loss functions.
We have utilized multiple colors to indicate different organs. Our investigation deals with the Bladder, Rectum, Sigmoid colon, and Femoral head, and these are represented by green, blue, red, and yellow colors, respectively. In addition, it should also be noted that the background will definitely come into existence while addressing the multiorgan segmentation task. Therefore, the background is represented by black. In this regard, Class 0, Class 1, Class 2, Class 3, and Class 4 are represented in the background (black), Bladder (green), Rectum (blue), Sigmoid colon (red), and Femoral head (yellow), respectively. The obtained results are presented in the following subsections.
Results on axial plane
This section presents detailed simulation results obtained under the proposed RU-Net architecture and the state-of-the-art methods. The detailed comparative analysis is presented considering different parameters, for example, different performance metrics and intersection over union (IOU) method. Therefore, the comparative analysis considering the performance metrics under CCE and dice loss is presented in Tables 1 and 2, respectively. All the results shown in the above tables are the mean values across the 5 classes in the multiorgan segmentation problem. In addition, a detailed result analysis is presented in Tables 3 and 4, considering the IOUs of the individual classes under CCE and dice loss, respectively.
Table 1.
Performance metrics considering categorical cross-entropy loss for axial plane
| Models | Accuracy | Precision | Recall | Specificity | F1-score | IOU |
|---|---|---|---|---|---|---|
| PSP-NET | 0.9989 | 0.8624 | 0.8638 | 0.9746 | 0.8619 | 0.7690 |
| ResU-Net | 0.9982 | 0.8132 | 0.8609 | 0.9737 | 0.8356 | 0.7303 |
| U-Net++ | 0.9986 | 0.8883 | 0.8655 | 0.9749 | 0.8810 | 0.7826 |
| U-Net | 0.9988 | 0.8710 | 0.9038 | 0.9827 | 0.8869 | 0.8051 |
| RU-Net (proposed) | 0.9986 | 0.8752 | 0.8717 | 0.9764 | 0.8727 | 0.7836 |
IOU: Intersection over union, PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net
Table 2.
Performance metrics considering dice loss for axial plane
| Models | Accuracy | Precision | Recall | Specificity | F1-score | IOU |
|---|---|---|---|---|---|---|
| PSP-NET | 0.9985 | 0.8748 | 0.8549 | 0.9710 | 0.8644 | 0.7726 |
| ResU-Net | 0.9986 | 0.8330 | 0.9007 | 0.9822 | 0.8649 | 0.7723 |
| U-Net++ | 0.9987 | 0.8693 | 0.9052 | 0.9807 | 0.8888 | 0.8041 |
| U-Net | 0.9989 | 0.8581 | 0.9161 | 0.9857 | 0.8857 | 0.8040 |
| RU-Net (proposed) | 0.9989 | 0.8686 | 0.9150 | 0.9845 | 0.8909 | 0.8118 |
IOU: Intersection over union, PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net
Table 3.
Class-wise intersection over union for axial plane under categorical cross-entropy loss
| Models | Class-1 | Class-2 | Class-3 | Class-4 | Class-5 |
|---|---|---|---|---|---|
| PSP-NET | 0.9965 | 0.8219 | 0.7105 | 0.5564 | 0.7598 |
| ResU-Net | 0.9957 | 0.7852 | 0.6773 | 0.5498 | 0.6438 |
| U-Net++ | 0.9968 | 0.8382 | 0.7340 | 0.5663 | 0.7778 |
| U-Net | 0.9972 | 0.8520 | 0.7624 | 0.6213 | 0.7924 |
| RU-Net (proposed) | 0.9966 | 0.8167 | 0.7297 | 0.5924 | 0.7826 |
IOU: Intersection over union, PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net
Table 4.
Class-wise intersection over union for axial plane under dice loss
| Models | Class-1 | Class-2 | Class-3 | Class-4 | Class-5 |
|---|---|---|---|---|---|
| PSP-NET | 0.9964 | 0.8229 | 0.7341 | 0.5562 | 0.7535 |
| ResU-Net | 0.9967 | 0.8299 | 0.7364 | 0.5820 | 0.7168 |
| U-Net++ | 0.9971 | 0.8390 | 0.7827 | 0.6209 | 0.7806 |
| U-Net | 0.9973 | 0.8562 | 0.7601 | 0.6052 | 0.8011 |
| RU-Net (proposed) | 0.9973 | 0.8572 | 0.7804 | 0.6161 | 0.8078 |
PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net
Tables 1-4 show the quantitative analysis for the axial plane, demonstrating the effectiveness of the proposed RU-Net architecture over PSP-Net, ResU-Net, U-Net++, and conventional U-Net. With dice-loss optimization [Table 2], RU-Net achieved the best overall IOU of 0.8118, which was higher than U-Net++ (0.8041), U-Net (0.8040), PSP-Net (0.7726), and ResU-Net (0.7723). It corresponds to an approximate improvement of 0.96% over the best competing baseline. Furthermore, the class-wise IOU analysis in Table 4 confirms that RU-Net gives better segmentation performance for difficult anatomical structures. In particular, RU-Net obtained the highest IOU of 0.8572 in Class-2 and 0.8078 in Class-5, and also performed better in Class-4 (0.6161) compared to U-Net (0.6052) and ResU-Net (0.5820). These results indicate that the incorporation of RRDB modules improves feature propagation and boundary delineation, leading to more accurate segmentation of anatomically sensitive organs in axial MR images.
Furthermore, a comparative analysis is presented in Figure 2, considering the five deep-learning networks, i.e. PSP-Net, ResU-Net, U-Net++, U-Net, and RU-Net (proposed) under the dice loss only. Because, in the presented tabular data, the overall result is improved under the dice loss condition. In order to support this, Figure 3 is utilized, where all the predictions are presented considering the proposed network and dice loss.
Figure 2.
Comparative assessment of the proposed RUNet architecture considering the axial plane
Figure 3.
Qualitative assessment of the proposed RUNet architecture under axial plane
Results on the coronal plane
This section exhibits detailed results considering CCE and dice loss conditions, under the five deep-learning networks, under the coronal plane. The performance metrics considering CCE and dice loss are presented in Tables 5 and 6, respectively.
Table 5.
Performance metrics considering categorical cross-entropy loss for coronal plane
| Models | Accuracy | Precision | Recall | Specificity | F1-score | IOU |
|---|---|---|---|---|---|---|
| PSP-NET | 0.9986 | 0.9490 | 0.9529 | 0.9901 | 0.9510 | 0.9080 |
| ResU-Net | 0.9984 | 0.9473 | 0.9474 | 0.9859 | 0.9449 | 0.8975 |
| U-Net++ | 0.9984 | 0.9406 | 0.9489 | 0.9885 | 0.9446 | 0.8971 |
| U-Net | 0.9987 | 0.9483 | 0.9593 | 0.9913 | 0.9544 | 0.9143 |
| RU-Net (Proposed) | 0.9985 | 0.9445 | 0.9561 | 0.9907 | 0.9503 | 0.9068 |
IOU: Intersection over union, PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net
Table 6.
Performance metrics considering dice loss for coronal plane
| Models | Accuracy | Precision | Recall | Specificity | F1-score | IOU |
|---|---|---|---|---|---|---|
| PSP-NET | 0.9980 | 0.9335 | 0.9346 | 0.9838 | 0.934 | 0.8787 |
| ResU-Net | 0.9977 | 0.9342 | 0.9319 | 0.9849 | 0.9330 | 0.8768 |
| U-Net++ | 0.9983 | 0.9431 | 0.9437 | 0.9872 | 0.9434 | 0.8947 |
| U-Net | 0.9981 | 0.9376 | 0.9456 | 0.9878 | 0.9429 | 0.8938 |
| RU-Net (proposed) | 0.9980 | 0.9322 | 0.9409 | 0.9862 | 0.9395 | 0.9163 |
IOU: Intersection over union, PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net
The detailed results are presented based on the mean values of the performance metric parameters and the IOUs obtained across the 5 classes. Furthermore, a detailed performance assessment is illustrated considering the class-wise IOUs under CCE and dice loss in Tables 7 and 8, respectively.
Table 7.
Classwise intersection over union for coronal plane under categorical cross-entropy loss
| Models | Class-1 | Class-2 | Class-3 | Class-4 | Class-5 |
|---|---|---|---|---|---|
| PSP-NET | 0.9964 | 0.9231 | 0.8596 | 0.8458 | 0.9152 |
| ResU-Net | 0.9960 | 0.9156 | 0.8448 | 0.821 | 0.9104 |
| U-Net++ | 0.9960 | 0.9137 | 0.8459 | 0.8208 | 0.9094 |
| U-Net | 0.9966 | 0.9265 | 0.8688 | 0.8563 | 0.9232 |
| RU-Net (proposed) | 0.9955 | 0.8907 | 0.8625 | 0.8257 | 0.9243 |
PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net
Table 8.
Classwise intersection over union for coronal plane under dice loss
| Models | Class-1 | Class-2 | Class-3 | Class-4 | Class-5 |
|---|---|---|---|---|---|
| PSP-NET | 0.9948 | 0.8877 | 0.8178 | 0.7992 | 0.8941 |
| ResU-Net | 0.9948 | 0.8790 | 0.8231 | 0.7966 | 0.887 |
| U-Net++ | 0.9956 | 0.9003 | 0.8472 | 0.8208 | 0.9094 |
| U-Net | 0.9955 | 0.8959 | 0.8345 | 0.8317 | 0.9117 |
| RU-Net (proposed) | 0.9955 | 0.915 | 0.8739 | 0.8685 | 0.9285 |
PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net
The effectiveness of the proposed RUNet architecture is further confirmed with the quantitative evaluation for the coronal plane, as shown in Tables 5-8. For dice-loss [Table 6], RU-Net achieved the best IOU of 0.9163, surpassing U-Net++ (0.8947), U-Net (0.8938), PSP-Net (0.8787), and ResU-Net (0.8768). The proposed model achieved improvements of 2.41% over U-Net++ and 2.52% over the conventional U-Net. Furthermore, the class-wise IOU analysis in Table 8 shows that RU-Net consistently achieves the best segmentation performance across several organ classes. Notably, RU-Net achieved the best Class-2 IOU of 0.9150, Class-3 IOU of 0.8739, Class-4 IOU of 0.8685, and Class-5 IOU of 0.9285, all significantly higher than those of the baseline methods. These findings indicate that the proposed RRDB-based architecture can effectively preserve structural information and improve organ boundary detection in coronal-plane MR images.
In connection with the earlier discussions, a comparative analysis considering the five deep-learning models is presented in Figure 4. Moreover, to conduct a detailed investigation, Figure 5 is presented, considering some other slices under the coronal plane in the presence of the proposed network and dice loss.
Figure 4.
Comparative assessment of the proposed RU-Net architecture considering the coronal plane
Figure 5.
Qualitative assessment of the proposed R-UNet architecture under the coronal plane
Results on the sagittal plane
This section presents detailed results for the different deep-learning configurations under CCE and dice loss, using image slices from the sagittal plane. The quantitative analysis considering the mean performance metrics and IOUs under CCE and dice loss is presented in Tables 9 and 10, respectively. The class-wise IOUs are presented considering the CCE and dice loss in Tables 11 and 12, respectively. The performance analysis in the sagittal plane also reveals the strong segmentation ability of the proposed RU-Net model, as shown in Tables 9-12. Considering CCE loss [Table 9], RU-Net achieved the best overall IOU of 0.9244, better than ResU-Net (0.9044), PSP-Net (0.8983), U-Net (0.8968), and U-Net++ (0.8938). This corresponds to an improvement of about 2.21% over ResU-Net and 3.08% over U-Net. Under dice-loss conditions [Table 10], likewise, RU-Net achieved the highest IOU of 0.9251, significantly outperforming all the competing methods. The class-wise analysis in Tables 11 and 12 further supports the robustness of the proposed method for complex anatomical structures. From Table 12, we can see that the Class-2 IOU of RU-Net is 0.9315, the Class-3 IOU is 0.9156, the Class-4 IOU is 0.8565, and the Class-5 IOU is 0.9266, which show consistent improvements over all baseline methods. These results confirm that the RRDB-enhanced RU-Net architecture effectively captures discriminative spatial features and enables superior multiorgan segmentation performance in sagittal MR images. Figures 6 and 7 present the comparative analysis between the five deep-learning models considering different slices of the sagittal plane under the application of dice loss.
Table 9.
Performance metrics considering categorical cross-entropy loss for sagittal plane
| Models | Accuracy | Precision | Recall | Specificity | F1-score | IOU |
|---|---|---|---|---|---|---|
| PSP-NET | 0.9982 | 0.9474 | 0.9435 | 0.9871 | 0.9455 | 0.8983 |
| ResU-Net | 0.9984 | 0.9463 | 0.9518 | 0.9895 | 0.9488 | 0.9044 |
| U-Net++ | 0.9982 | 0.9331 | 0.9534 | 0.9889 | 0.9428 | 0.8938 |
| U-Net | 0.9982 | 0.9203 | 0.9708 | 0.9932 | 0.9445 | 0.8968 |
| RU-Net (Proposed) | 0.9987 | 0.9614 | 0.9592 | 0.9920 | 0.9602 | 0.9244 |
PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net, IOU: Intersection over union
Table 10.
Performance metrics considering dice loss for sagittal plane
| Models | Accuracy | Precision | Recall | Specificity | F1-score | IOU |
|---|---|---|---|---|---|---|
| PSP-NET | 0.9979 | 0.9370 | 0.9361 | 0.9855 | 0.9365 | 0.8827 |
| ResU-Net | 0.9979 | 0.9364 | 0.9389 | 0.9865 | 0.9376 | 0.8848 |
| U-Net++ | 0.9982 | 0.9411 | 0.9502 | 0.9888 | 0.9456 | 0.8984 |
| U-Net | 0.9982 | 0.9458 | 0.9481 | 0.9884 | 0.9469 | 0.9008 |
| RU-Net (Proposed) | 0.9981 | 0.9366 | 0.9503 | 0.9887 | 0.9433 | 0.9251 |
PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net, IOU: Intersection over union
Table 11.
Classwise intersection over union for sagittal plane under categorical cross-entropy loss
| Models | Class-1 | Class-2 | Class-3 | Class-4 | Class-5 |
|---|---|---|---|---|---|
| PSP-NET | 0.9954 | 0.9159 | 0.8627 | 0.8264 | 0.8911 |
| ResU-Net | 0.9966 | 0.9389 | 0.8769 | 0.8732 | 0.9208 |
| U-Net++ | 0.9953 | 0.9180 | 0.8638 | 0.8067 | 0.8853 |
| U-Net | 0.9956 | 0.9228 | 0.8510 | 0.8203 | 0.8941 |
| RU-Net (proposed) | 0.9967 | 0.9435 | 0.8825 | 0.8755 | 0.9238 |
PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net
Table 12.
Classwise intersection over union for sagittal plane under dice loss
| Models | Class-1 | Class-2 | Class-3 | Class-4 | Class-5 |
|---|---|---|---|---|---|
| PSP-NET | 0.9945 | 0.8936 | 0.8442 | 0.8017 | 0.8796 |
| ResU-Net | 0.9947 | 0.9012 | 0.8468 | 0.7971 | 0.8841 |
| U-Net++ | 0.9954 | 0.9103 | 0.8667 | 0.8253 | 0.8944 |
| U-Net | 0.9955 | 0.9153 | 0.8688 | 0.8256 | 0.8988 |
| RU-Net (proposed) | 0.9952 | 0.9315 | 0.9156 | 0.8565 | 0.9266 |
PSP-NET: Pyramid scene parsing network, ResU-Net: Residual U-Net, RU-Net: Residual in Residual Dense Block U-Net
Figure 6.
Comparative assessment of the proposed RU-Net architecture considering the sagittal plane
Figure 7.
Qualitative assessment of the proposed RU-Net architecture under the sagittal plane
DISCUSSION
This section includes a detailed discussion of the results presented in Section 3 for all three planes, considering both tabular and image data. According to the performance metrics analysis shown in Tables 1, 2, 5, 6a, 9, and 10, it can be stated that the proposed architecture shows superiority in performance compared to the other state-of-the-art methods when tested under both the CCE and dice loss scenarios. A range of 0.1%−0.59% improvement has been observed in IOU, while the best performance in terms of IOU has been found in the sagittal plane. Based on the overall results across all planes, it can be stated that a significant improvement is achieved with the proposed network under the dice loss scenario compared to conventional methods.
Based on the class-wise detailed results analysis, presented in Tables 3, 4, 7, 8, 11, 12, it can be noticed that the proposed RU-Net shows superior results compared to the other tested methods. These class-wise results can be considered a continuation to the performance metrics analysis. In addition, critical observations indicate that class 3 and class 4 are very sensitive and present lower IOUs compared to the other classes. A significant improvement is achieved under the proposed RU-Net network with both losses; while the best improvement compared to U-Net for class 4 has been found in the axial plane (2.14%), and for class 5, it was observed in the sagittal plane (3.82%). Analyzing all the obtained results, it can be established that the best performance can be assured considering the proposed architecture under the dice loss condition. Moreover, to rigorously evaluate the generalization performance and training stability of the proposed RU-Net framework against baseline architectures, we have performed a comprehensive statistical analysis based on multiplane cross-validation metrics. We conducted rigorous experiments on the test dataset to confirm the stability and precision of the models, evaluating out-of-distribution robustness and clinical generalization capability. The results demonstrate that RU-Net achieved the highest mean IOU values and the lowest standard deviation across axial (±5.24%), coronal (±5.49%), and sagittal (±4.77%) views, indicating superior and stable segmentation performance. Considering the average across all planes, U-Net (±6.25%) and U-Net++ (±7.04%) also exhibited strong performance, with relatively narrow distributions and comparable variances, reflecting stable predictions. In contrast, PSP-Net produced the lowest mean IOU values, particularly in the axial plane, while Residual U-Net showed the greatest variability across cases. These findings indicate that recurrent residual feature extraction in RU-Net improves segmentation robustness and accuracy compared with the other evaluated architectures.
On critical observation, we find that for the coronal plane [Table 5 under CCE loss, and Table 6 considering the IOUs for dice loss], the proposed network and the U-Net network have comparable performance. Therefore, in these cases, it is also important to analyze the qualitative results visually by expert radiologists. According to the image-based results depicted in Figures 2-7, it is confirmed that the overall improved results are obtained considering the Proposed RU-Net under dice loss condition. The edges are efficiently detected using the proposed network. It should also be stated that every trained model has some shortcomings. Therefore, a misclassified image is also presented in Figure 3 to justify the proposed network under the axial plane. In addition, it is also noticed that, although for the coronal plane, the IOU is slightly better in case of U-Net-dice compared to the proposed RU-Net-dice, based on the comparative analysis shown in Figure 4, the superior performance of the proposed RU-Net-dice can be established in terms of proper prediction and efficient edge detection of the classes, mainly the sensitive classes.
Based on the detailed discussion, it is established that the proposed RU-Net shows its supremacy under dice loss conditions in terms of proper prediction and segmentation from the available ground-truth data.
CONCLUSION
BT stands at the forefront of cancer treatment modalities, offering precise and localized radiation therapy that maximizes tumor control while minimizing damage to surrounding healthy tissues. Central to the success of BT is the accurate delineation of target volumes and OARs during treatment planning, ensuring optimal dose delivery and minimizing potential toxicity. Among the critical steps in this process, multiorgan image segmentation plays a pivotal role, yet remains a formidable challenge due to the inherent complexity and variability of anatomical structures. Current segmentation techniques, although diverse in approach, often struggle to achieve the level of accuracy and efficiency demanded by modern BT practice. Challenges related to BT probe-related deformation of the OARs, interpatient anatomical variation, image artifacts, and the proximity of organs pose a significant hurdle to reliable segmentation, potentially leading to suboptimal treatment outcomes and increased risk of complications.
In this study, we have utilized a clinical MR image dataset comprising 95 patients who underwent intracavitary BT dataset to assess the proposed RRDB-based method for automatic segmentation of bladder, rectum, sigmoid colon, and femoral head in high-dose rate BT. Before this work, no extensive studies have been reported for MR images with the Fletcher-Suit applicator. The proposed model was trained and tested for all three planes (i.e., axial, coronal, and sagittal) with this substantial patient dataset. The results indicate that the proposed model can extract more information effectively, thanks to the inclusion of RRDB layers. The results also indicate that the proposed network, when using dice loss, achieves the highest IOUs across different image frames. Thus, we report that integrating the RRDB model enhances precision, efficiency, and effectiveness in multiorgan segmentation for BT, ultimately leading to better patient outcomes. Using this method, accurate segmentation, precise dose optimization, better tumor characterization, and personalized treatment plans can be tailored to individual patient anatomy. By reducing the risk of underdosing target volumes or irradiating healthy tissues, our proposed methodology has the potential to elevate the standard of care in BT, leading to improved local control rates, reduced toxicity, and enhanced quality of life for cancer patients.
Data availability
Data will be available with the corresponding author on reasonable request.
Conflicts of interest
There are no conflicts of interest.
Funding Statement
Nil.
REFERENCES
- 1.Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68:394–424. doi: 10.3322/caac.21492. [DOI] [PubMed] [Google Scholar]
- 2.Global Burden of Disease Cancer Collaboration. Fitzmaurice C, Dicker D, Pain A, Hamavid H, Moradi-Lakeh M, et al. The global burden of cancer 2013. JAMA Oncol. 2015;1:505–27. doi: 10.1001/jamaoncol.2015.0735. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Abu-Rustum, Susana M Campos, Sudha Amarnath, Rebecca Arend, Emma Barber, Kristin Bradley et al. “NCCN guidelines® insights: cervical cancer, version 1.2024: featured updates to the NCCN guidelines.” Journal of the National Comprehensive Cancer Network 21.12 2023. p. 1224-233. [[Last accessed on 2026 Jul 19]]. Available from: http://www. NCCN.org/professionals/physician gls/PDF/occult.Pdf . [DOI] [PubMed]
- 4.Ahn SH, Yeo AU, Kim KH, Kim C, Goh Y, Cho S, et al. Comparative clinical evaluation of atlas and deep-learning-based auto-segmentation of organ structures in liver cancer. Radiat Oncol. 2019;14:213. doi: 10.1186/s13014-019-1392-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Major T, Fröhlich G, Ágoston P, Polgár C, Takácsi-Nagy Z. The value of brachytherapy in the age of advanced external beam radiotherapy: A review of the literature in terms of dosimetry. Strahlenther Onkol. 2022;198:93–109. doi: 10.1007/s00066-021-01867-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Schroeder S, Tan J, Yang M, Tian Z, Jia X, Albuquerque K. Clinical implementation of automated ICRU 89 reporting for definitive management of cervical cancer with intracavitary brachytherapy. Brachytherapy. 2018;17:S112–3. [Google Scholar]
- 7.Feng X, Chen Q. Artificial intelligence for image segmentation in radiation oncology. Artificial Intelligence in Radiation Oncology. AME Publishing Company: World Scientific. 2023:103–28. [Google Scholar]
- 8.Kiser KJ, Fuller CD, Reed VK. Artificial Intelligence in Radiation Oncology Treatment Planning: A Brief Overview. J Med Artificial Intelligence. 2019:2. [Google Scholar]
- 9.Li C, Mao Y, Guo Y, Li J, Wang Y. Multi-dimensional cascaded net with uncertain probability reduction for abdominal multi-organ segmentation in CT sequences. Comput Methods Programs Biomed. 2022;221:106887. doi: 10.1016/j.cmpb.2022.106887. [DOI] [PubMed] [Google Scholar]
- 10.Li X, Hu Y. Cooperative-net: An end-to-end multi-task interaction network for unified reconstruction and segmentation of MR image. Comput Methods Programs Biomed. 2024;245:108045. doi: 10.1016/j.cmpb.2024.108045. [DOI] [PubMed] [Google Scholar]
- 11.Benčević M, Habijan M, Galić I, Babin D, Pižurica A. Understanding skin color bias in deep learning-based skin lesion segmentation. Comput Methods Programs Biomed. 2024;245:108044. doi: 10.1016/j.cmpb.2024.108044. [DOI] [PubMed] [Google Scholar]
- 12.Cardenas CE, Yang J, Anderson BM, Court LE, Brock KB. Advances in auto-segmentation. In Seminars in radiation oncology. WB Saunders: Elsevier. 2019;29:185–97. doi: 10.1016/j.semradonc.2019.02.001. [DOI] [PubMed] [Google Scholar]
- 13.Basu A, Senapati P, Deb M, Rai R, Dhal KG. A survey on recent trends in deep learning for nucleus segmentation from histopathology images. Evol Syst. 2024;15:203–48. doi: 10.1007/s12530-023-09491-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Kim N, Chang JS, Kim YB, Kim JS. Atlas-based auto-segmentation for postoperative radiotherapy planning in endometrial and cervical cancers. Radiat Oncol. 2020;15:106. doi: 10.1186/s13014-020-01562-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Teguh DN, Levendag PC, Voet PW, Al-Mamgani A, Han X, Wolf TK, et al. Clinical validation of atlas-based auto-segmentation of multiple target volumes and normal tissue (swallowing/mastication) structures in the head and neck. Int J Radiat Oncol Biol Phys. 2011;81:950–7. doi: 10.1016/j.ijrobp.2010.07.009. [DOI] [PubMed] [Google Scholar]
- 16.Kusada T, Toita T, Ariga T, Maemoto H, Hashimoto S, Shiina H, et al. Computed tomography–based image-guided brachytherapy for cervical cancer: Correlations between dose– volume parameters and clinical outcomes. J Radiat Res. 2018;59:67–76. doi: 10.1093/jrr/rrx065. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Chi Y, Pan Y, Zhang N, Han D, Guo X, Mao Z, et al. Clinical outcomes of MRI-guided adaptive brachytherapy for each fraction in locally advanced cervical cancer: A single institution experience. Front Oncol. 2022;12:841980. doi: 10.3389/fonc.2022.841980. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Koutoulakis E, Marage L, Markodimitrakis E, Aubignac L, Jenny C, Bessieres I, et al. Automatic multiorgan segmentation in pelvic region with convolutional neural networks on 0.35 tmr-linac images. Algorithms. 2023;16:521. [Google Scholar]
- 19.Kalantar R, Lin G, Winfield JM, Messiou C, Lalondrelle S, Blackledge MD, et al. Automatic segmentation of pelvic cancers using deep learning: State-of-the-Art approaches and challenges. Diagnostics (Basel) 2021;11:1964. doi: 10.3390/diagnostics11111964. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Zhu Y, Wei R, Gao G, Ding L, Zhang X, Wang X, et al. Fully automatic segmentation on prostate MR images based on cascaded fully convolution network. J Magn Reson Imaging. 2019;49:1149–56. doi: 10.1002/jmri.26337. [DOI] [PubMed] [Google Scholar]
- 21.Li Z, Chen K, Yang Z, Zhu Q, Yang X, Li Z, et al. A personalized DVH prediction model for HDR brachytherapy in cervical cancer treatment. Front Oncol. 2022;12:967436. doi: 10.3389/fonc.2022.967436. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Wang J, Chen Y, Tu Y, Xie H, Chen Y, Luo L, Zhou P, Tang Q. Evaluation of auto-segmentation for brachytherapy of postoperative cervical cancer using deep learning-based workflow. Physics in Med Biol. 2023;68:055012. doi: 10.1088/1361-6560/acba76. [DOI] [PubMed] [Google Scholar]
- 23.Comelli A, Dahiya N, Stefano A, Vernuccio F, Portoghese M, Cutaia G, et al. Deep learning-based methods for prostate segmentation in magnetic resonance imaging. Appl Sci (Basel) 2021;11:782. doi: 10.3390/app11020782. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Li Z, Zhu Q, Zhang L, Yang X, Li Z, Fu J. A deep learning-based self-adapting ensemble method for segmentation in gynecological brachytherapy. Radiat Oncol. 2022;17:152. doi: 10.1186/s13014-022-02121-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Hu D, Jian J, Li Y, Gao X. Deep learning-based segmentation of epithelial ovarian cancer on T2-weighted magnetic resonance images. Quant Imaging Med Surg. 2023;13:1464–77. doi: 10.21037/qims-22-494. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Ni R, Han K, Haibe-Kains B, Rink A. Generalizability of deep learning in organ-at-risk segmentation: A transfer learning study in cervical brachytherapy. Radiother Oncol. 2024;197:110332. doi: 10.1016/j.radonc.2024.110332. [DOI] [PubMed] [Google Scholar]
- 27.Fransson S. Comparing multi-image and image augmentation strategies for deep learning-based prostate segmentation. Phys Imaging Radiat Oncol. 2024;29:100551. doi: 10.1016/j.phro.2024.100551. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Das S, Sanki P, Mondal S, Bhattacharya S, Ghosh S. Bladder segmentation in MRI for high dose rate brachytherapy using deep network. Singapore: Springer Nature Singapore. In International Conference on Emerging Applications of Information Technology 2024. :55–67. [Google Scholar]
- 29.Zhou Z, Siddiquee MM, Tajbakhsh N, Liang J. UNet++: A nested U-Net architecture for medical image segmentation. Deep Learn Med Image Anal Multimodal Learn Clin Decis Support (2018) 2018;11045:3–11. doi: 10.1007/978-3-030-00889-5_1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Ronneberger O, Fischer P, Brox T. U net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer. 2015:234–41. [Google Scholar]
- 31.Dai H, Xie W, Xia E. SK-Unet++: An improved Unet++ network with adaptive receptive fields for automatic segmentation of ultrasound thyroid nodule images. Med Phys. 2024;51:1798–811. doi: 10.1002/mp.16672. [DOI] [PubMed] [Google Scholar]
- 32.Zhao H, Shi J, Qi X, Wang X, Jia J. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017:2881–90. [Google Scholar]
- 33.Diakogiannis FI, Waldner F, Caccetta P, Wu C. Resunet-a: A deep learning framework for semantic segmentation of remotely sensed data. ISPRS J Photogramm Remote Sens. 2020;162:94–114. [Google Scholar]
- 34.Alom MZ, Yakopcic C, Hasan M, Taha TM, Asari VK. Recurrent residual U-Net for medical image segmentation. J Med Imaging (Bellingham) 2019;6:014006. doi: 10.1117/1.JMI.6.1.014006. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Zhang Z, Sabuncu MR. Generalized cross entropy loss for training deep neural networks with noisy labels. Adv Neural Inf Process Syst. 2018;32:8792–802. [PMC free article] [PubMed] [Google Scholar]
- 36.Li X, Sun X, Meng Y, Liang J, Wu F, Li J. Dice loss for data-imbalanced NLP tasks. In Proceedings of the 58th annual meeting of the association for computational linguistics. 2020:465–76. [Google Scholar]
- 37.Zhao R, Qian B, Zhang X, Li Y, Wei R, Liu Y, et al. Rethinking Dice Loss for Medical Image Segmentation. In 2020 IEEE International Conference on Data Mining (ICDM) IEEE. 2020:851–60. [Google Scholar]
- 38.Manna S, Das S, Ghosh S, Bhattacharya S. Federated self-supervised learning for one-shot cross-modal and cross-imaging technique segmentation. arXiv preprint arXiv:2503.23507.2025 [Google Scholar]
- 39.Deora P, Vasudeva B, Bhattacharya S, Pradhan PM. Structure preserving compressive sensing MRI reconstruction using generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2020:522–3. [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
Data will be available with the corresponding author on reasonable request.