A modified deep learning approach for seminal vesicle region localization in prostate MRI

Abstract

The seminal vesicle region plays a crucial role in male reproductive health, and its accurate evaluation is essential for diagnosing infertility and carcinoma. Magnetic resonance imaging (MRI) is the primary modality for assessment; however, manual evaluation is time-consuming and subject to interobserver variability, necessitating automated approaches. This study presents a modified ResNet-based deep learning model specifically developed for automated localization of the seminal vesicle region in prostate MRI scans. Unlike segmentation-based methods, the focus is on robust region-level localization as a precursor to detailed analysis. To the best of our knowledge, this is the first study to address seminal vesicle localization directly from MRI using a deep learning model. Performance was evaluated using classification accuracy, inference time, and true positive (TP) coverage, with a sliding window approach to detect high-confidence regions. The modified ResNet-34 achieved the highest TP coverage (0.885) and classification accuracy (0.979), demonstrating improved localization with minimal computational overhead. Heatmap visualizations confirmed the model’s focus on relevant anatomical structures. The proposed approach provides a practical solution for reducing manual effort and interobserver variability, offering a reliable foundation for subsequent segmentation and abnormality detection. Future work may explore the integration of multi-view imaging or 3D CNN architectures to further improve performance.

Introduction

Seminal vesicles (also referred to as seminal or vesicular glands) are integral components or accessory glands of the male reproductive system, primarily responsible for producing seminal fluid, a major component of semen. The seminal vesicle region is typically located in the retrovesical space, posterior to the urinary bladder and anterior to the rectum¹.

Various pathological conditions may manifest within the seminal vesicles, including infections (vesiculitis or spermatocystitis), cysts, and tumor invasion. MRI is a valuable imaging modality for the detailed evaluation of seminal vesicle abnormalities and treatment planning. T2-weighted images best view the low-signal intensity capsule and the seminal vesicle convolutions². Consequently, accurate manual delineation of this region on MRI data is a time-consuming process and a critical task for expert radiologists, playing an essential role in adaptive radiotherapy.

Subjectivity and the labor-intense nature of the inspection creates a strategic demand for automation³.To the best of our knowledge, no previous work has addressed direct MRI-based localization of the seminal vesicle using deep learning, especially with a ResNet architecture focused on coronal T2 slices. However emerging research suggests that computer-aided localization of prostate-related structures in MRI can significantly enhance accuracy and speed while simultaneously optimizing the time required for analysis⁴. In line with this, a recent systematic review and meta-analysis by Zhao et al. confirmed the strong diagnostic value of machine learning models in prostate MRI, achieving pooled AUC values up to 0.96. These findings indicate the clinical potential of AI-based methods and further justify the application of deep learning to adjacent anatomical structures such as the seminal vesicles⁵. Deep learning approaches are among the most widely and effectively adopted machine learning techniques for accurate automated localization and segmentation of prostate-related regions in MRI data. The multi-layered structure of deep learning-based supervised methods provides a high-level representation of the input data for improving the predictions⁶. Accordingly, various deep-learning models have previously been adopted for prostate-related segmentation tasks. ResNet is one of the most widely adopted deep learning models for prostate region analysis⁷. Furthermore, ResNet has shown strong performance across different anatomical regions, making it a competitive alternative for automated segmentation tasks. Accordingly, a modified ResNet architecture is implemented in this study to automate the recognition of the targeted area. A key objective of this research is to modify and train the model to specialize in precisely identifying the seminal vesicle region in MRI data. Since the seminal vesicle is located within commonly targeted areas, several studies have explored the use of deep learning for delineating or detecting regions that include it. For instance, Takeshita et al. developed a CNN-based model for recognizing the seminal vesicle and vas deferens during robot-assisted radical prostatectomy, achieving a Dice similarity coefficient of 0.73. However, their approach focuses on intraoperative video-based segmentation rather than MRI-based analysis⁸.

Although many studies have focused on prostate segmentation and classification, very few have directly examined the seminal vesicle, which underscores the novelty of our approach. This research aims to develop an effective automated mechanism for accurate region identification by implementing a modified ResNet architecture. A heatmap-based localization mechanism was implemented to qualitatively assess the model’s focus on the correct anatomical structure. The resulting heatmaps provide a probabilistic spatial distribution of the detected region and offer an intuitive representation of the model’s decision-making process.

Literature review

Although deep learning has been widely adopted for the segmentation of specific regions in MRI, only a limited number of studies have directly examined the seminal vesicle region, indicating a significant gap in the literature. Nevertheless, extensive research has been carried out on prostate MRI, where deep learning methods are now widely used for automated analysis, both for segmentation and classification. In one of the early research studies, Barra et al. implemented the U-Net model for automatic prostate segmentation on bicentric T2-weighted images. They have reported that their implementation can segment the region with a Dice similarity coefficient ⁹. Furthermore, a fully automated deep learning system developed by Pellicer-Valero et al. was trained for prostate cancer detection and achieved an AUC of up to 0.96 for lesion-level classification¹⁰. Similarly, a 3D convolutional neural network (CNN) achieved an accuracy of in detecting prostate cancer from MRI scans¹¹. In addition to these advances, the U-Net framework introduced by Ronneberger et al. has become a standard model for medical image segmentation. It has been widely applied in prostate imaging studies¹². Later studies introduced CNN-based methods. For instance, a recent study proposed a deep learning framework that combines CNN-based cancer detection with image post-processing to localize extraprostatic extension (EPE) in prostate MRI, and the model reportedly outperformed radiologists in sensitivity¹³. Another recent work presented RACLAHE, a novel image enhancement method to improve CNN-based prostate and zonal segmentation on MRI. This approach was tested across multiple architectures and consistently improved Dice scores¹⁴. More recent works combined CNNs with attention or transformer modules. For example, Cai et al. introduced a deep Transformer-based framework implemented for prostate MRI segmentation. By integrating a Deep Union Transformer encoder and a Pool Fusion Attention decoder, their method effectively captured both global and local features, outperforming state-of-the-art approaches across multiple datasets.¹⁵. ResNet has been one of the most widely used backbones in prostate MRI segmentation studies. The systematic review by Fassia et al. reported that deep learning architectures, including ResNet, have achieved segmentation performance that, in several studies, was comparable to expert radiologists⁷. Similarly, a recent review by Valizadeh et al. highlighted the role of ResNet, DenseNet, and EfficientNet in prostate cancer classification from multiparametric MRI¹⁶. In summary, deep learning has become a well-established tool for prostate segmentation and classification, but direct localization of the seminal vesicle region in MRI has rarely been studied. This creates a clear gap that our study aims to address with a modified ResNet model trained on T2-weighted images.

Material and methods

Original data

This retrospective study investigated the patients who underwent prostate MRI in the Izmir City Hospital. The study was conducted in accordance with the Declaration of Helsinki, and approved by the Institutional Review Board of Izmir City Hospital (approval number: 2025/49, Date: 05/02/2025). Informed consent was obtained from all subjects involved in the study. Two hundred and one coronal T2-weighted MR images were collected from Izmir City Hospital to form the original dataset. All patients were between the ages of 40-89. Patients with poor image quality and those who did not refrain from ejaculation for 3 days or more before MRI were excluded from the study to improve the diagnostic interpretation of SVs¹⁷. An expert radiologist manually annotated the MRI images to establish the ground truth. Of these, 140 images were used for training, 30 for validation, and 31 for testing.

Preprocessing

The original MRI images varied in size and were not standardized. The average resolution across the dataset was approximately 1372 1249 pixels. All MRI images were first converted to grayscale and resized to one-third of their original dimensions using the area interpolation method, which effectively downsamples images to reduce computational load while preserving essential anatomical details. Additionally, MRI scans were converted into a three-channel format to ensure compatibility with the ResNet architecture, allowing the use of pre-trained weights. This preprocessing ensured consistent resolution across all images, reduced computational complexity, and optimized training time. After resizing, a manual region selection process was applied to the training and validation images, where an expert interactively extracted 50x50 pixel sub-regions by selecting relevant areas. As a result, 517 negative and 517 positive equally sized images were saved for training, while 115 negative and 115 positive images were designated for validation. Positive and negative image samples refer to regions extracted from the original MRI images that do or do not contain any part of the seminal vesicles, respectively. The labeling was performed manually by an expert radiologist based on visual identification of the vesicle structures. The preprocessing stages are illustrated in Fig. 1 using a sample region.

Fig. 1.

Illustration of the preprocessing stages on a sample region.

Deep learning based approach

A ResNet-based modified model was implemented in this study. ResNet (Residual Neural Network) is a deep learning architecture designed by integrating residual connections into conventional deep networks¹⁸. These residual units help address challenges associated with training deep neural networks¹⁹. A typical residual unit consists of a stack of convolutional layers, batch normalization, and non-linear activation functions, which collectively mitigate the vanishing/exploding gradient problem²⁰.

A typical residual unit begins with a convolutional layer for extracting spatial features, followed by batch normalization to stabilize learning. A ReLU activation function is then applied, allowing the network to learn complex patterns. Another convolutional layer and a second batch normalization step for further refinement follow this. The defining feature is the skip (or identity) connection, which bypasses the intermediate transformations and adds the input x directly to the output of the layer stack F(x), forming the residual mapping:

1

Here, F(x) represents the residual function (typically a sequence of Conv-BN-ReLU-Conv-BN), and x is the input passed unchanged through the identity path. This connection helps preserve essential information while ensuring gradients flow effectively, preventing vanishing or exploding issues. Fig. 2 illustrates the original residual block components.

Fig. 2.

Structure of a basic residual block.

Modified model architecture

Two ResNet topologies are evaluated in this study: ResNet-18 and ResNet-34. ResNet-18 was selected due to its faster training time and lower number of parameters, while ResNet-34 was included to investigate the impact of increased model depth. The fully connected (FC) layer was modified for binary classification (positive and negative regions) to adapt the model for the classification task. A redesigned, fully connected block was also introduced instead of the original final layer. This alternative block consists of a linear layer with 256 neurons, Batch Normalization to stabilize learning, and ReLU activation for non-linearity, enhancing the model’s learning capacity and stability for the primary classification task. As the final output layer, a second linear layer maps the features to two output neurons for binary classification. The overall structure of the modified model is illustrated in Fig. 3.

Fig. 3.

Comparison of the original and modified ResNet architectures.

Training process

The modified architecture was trained fully supervised with the cross-entropy loss function to optimize classification performance. The AdamW optimizer was applied with a learning rate 0.0001 and a weight decay of 1e-4, ensuring controlled weight updates to enhance training stability and convergence. A learning rate scheduler was also implemented to reduce the learning rate by 0.1 every five epochs. The model was trained for 10 epochs using a batch size of 32. Images were normalized using ImageNet statistics to align with the model’s pre-trained weights. It should be noted that all network layers were kept trainable to allow fine-tuning of feature representations specific to the seminal vesicle region. The train and validation loss curves for modified ResNet-18 and modified ResNet-34 models over 10 epochs are represented in Fig. 4.

Fig. 4.

Train-validation loss curves over 10 epochs. (a) Modified ResNet-18. (b) Modified ResNet-34.

Performance evaluation

Trained net is performed for locating the spatial location of the seminal vesicle region. Accordingly, test images are processed using a sliding window approach with 50x50 pixel patches and a 25-pixel step size. Each patch was classified using the trained model, and regions with a confidence score above 0.80 were selected as estimated positive windows. Two different criteria were calculated for numerical performance evaluation.

The first criterion is true positive (TP) coverage, which measures the proportion of the ground truth region that is covered by positive predictions. Higher TP coverage indicates better localization performance. The second criterion is classification accuracy, computed based on the true positive (TP), true negative (TN), false positive (FP), and false negative (FN) counts across the sliding window patches. Accuracy is calculated as:

2

The accuracy metric quantifies how well the model distinguishes between the positive and negative regions. Additionally, estimation time per image was recorded to assess the computational efficiency of the approach. Shorter inference times are preferred for real-time applications.

Results

Experimental setup

All the experiments were conducted on a computer with an Intel Core i7-9750H CPU running at 2.60 GHz, 16 GB of DDR4 RAM, and an NVIDIA GeForce GTX 1050 GPU. All experiments, including the training and testing of deep learning models, were performed using Python 3.9.4.

Model performance

This study evaluated two modified ResNet-based models. Additionally, DenseNet was trained and tested using the same experimental setup and parameters for a fair and objective comparison. DenseNet is one of the most widely utilized deep-learning models for prostate region segmentation.

Model performance was assessed using True Positive (TP) Coverage, classification accuracy, and inference time. TP Coverage quantifies the proportion of the ground truth region the model correctly detects. Accuracy was computed by evaluating patch-based classification results, where each 50x50 pixel window was classified as containing or not containing the seminal vesicle.

The inference time was measured as the total processing time required to analyze a single MRI slice, including patch extraction and model evaluation. The summarized results, including mean, standard deviation, maximum, and minimum values for TP coverage, classification accuracy, and inference time, are presented in Table 1. These values represent per-image evaluation results. For each test image, TP coverage, accuracy, and inference time were calculated individually, and the table presents the mean, standard deviation, minimum, and maximum values across all test images.

Table 1.

Performance comparison of original ResNet-18, modified ResNet-18, modified ResNet-34, and DenseNet models.

Original ResNet-18
Metric	Mean	Std	Max	Min
TP coverage	0.867	0.151	1.000	0.465
Accuracy	0.975	0.010	0.992	0.946
Time	1.423	0.038	1.521	1.379
Modified ResNet-18
Metric	Mean	Std	Max	Min
TP coverage	0.815	0.190	1.000	0.325
Accuracy	0.976	0.013	0.996	0.950
Time	1.461	0.049	1.613	1.379
Modified ResNet-34
Metric	Mean	Std	Max	Min
TP coverage	0.885	0.111	1.000	0.547
Accuracy	0.979	0.011	0.996	0.958
Time	2.427	0.058	2.552	2.339
Original DenseNet
Metric	Mean	Std	Max	Min
TP coverage	0.808	0.162	1.000	0.248
Accuracy	0.972	0.011	0.992	0.942
Time	4.055	0.077	4.207	3.924

It should be noted that, the ResNet model outputs a probability score for each 5050 pixel window, representing the likelihood that the window contains part of the seminal vesicle. A window is classified as positive if its predicted probability exceeds a threshold value of 0.80. This value was empirically determined to provide a good balance between sensitivity and specificity in detecting the seminal vesicle regions. This threshold-based decision determines the classification outcome (positive or negative) for each window, and is used to compute the corresponding counts of True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN). These window-level decisions are then used to calculate image-level TP Coverage and classification accuracy, as summarized in Table 1.

Comparison of the detection performance of the original ResNet-18 and modified ResNet-34 models on three sample test images are given in Fig. 5. Green regions indicate the ground truth of the seminal vesicle region, while blue regions represent the positive windows detected by the models. Each image includes both models’ TP coverage, classification accuracy, and processing time.

Fig. 5.

Detection performance of the original ResNet-18 and modified ResNet-34 is demonstrated on six sample images from the test data. (a) Original ResNet-18. (b) Modified ResNet-34. Green regions indicate the ground truth, while the blue boxes represent positive windows selected by the model.

Heatmap visuals

To further validate the accuracy of our model’s localization, heatmap-derived segmentation masks were generated and visually analyzed. These heatmaps provide a probabilistic spatial representation of the seminal vesicle region, allowing an intuitive interpretation of the model’s decision-making process. This approach enables more granular and interpretable localization by assigning a probability score to each small region within the image rather than making a binary classification of the entire MRI slice.

The heatmaps highlight high-confidence areas where the seminal vesicle is likely present, facilitating the creation of a refined segmentation mask by aggregating predictions across multiple regions. To achieve this, each MRI slice was divided into overlapping patches using a fixed window size of 50x50 pixels with a step size of 25 pixels. Each extracted patch was resized, normalized, and passed through the modified ResNet-34 model, which assigned a probability score for the presence of the seminal vesicle. These scores were then combined into a spatial heatmap, where each region’s intensity reflects the likelihood of the seminal vesicle being present.

It should be noted that during the classification process, the model assigns probability scores not to individual pixels, but to 5050 image regions spaced every 25 pixels. Consequently, the resulting heatmaps represent overlapping region-level predictions rather than pixel-wise classifications.

Fig. 6 presents examples of the generated heatmaps and their corresponding segmentation masks. The visual analysis confirms that the model consistently assigns high probability values to the expected anatomical region, reinforcing confidence in its localization ability.

Fig. 6.

Examples of generated heatmaps and corresponding segmentation masks for seminal vesicle localization. (a) Original MRI image. (b) Heatmap indicating high-probability regions. (c) Binary segmentation mask with TP coverage and inference time.

Discussion

The involvement of one or both seminal vesicles by prostate cancer is known as T3b disease, and the incidence of SV invasion in surgical series has been reported to range from 4 to 23%^21,22. Preoperative detection of SV involvement is crucial for prognosis and treatment planning because patients with T3b disease are more likely to experience lymph node involvement, local recurrence, and distant metastases²³. There are some studies about seminal vesicle involvement by prostate cancer using T2 signal intensity on MRI.^21,24,25. Seminal vesicle invasion is not always easy to recognize on MRI images and is typically identified as a low-signal area extending into the seminal vesicle on T2-weighted images. Due to disrupting the typical convolutional structure of the seminal vesicles, methods developed in this manner can facilitate the recognition of seminal vesicle tumor invasion. To address these challenges and improve the accuracy of seminal vesicle localization, this study evaluated the performance of original and modified ResNet-18, ResNet-34, and DenseNet models. True positive (TP) coverage, classification accuracy, and inference time were key evaluation metrics. The results indicate that the applied structural modifications have influenced accuracy, localization performance, and computational efficiency.

Compared to segmentation-based approaches such as those proposed by Pellicer-Valero et al.¹⁰, our method focuses on accurate localization as an initial step, achieving over 0.9 TP coverage and 97% classification accuracy. While segmentation models provide detailed boundary delineation, they often require higher data granularity and computational resources. Our approach, in contrast, offers a computationally efficient localization step, which can serve as a foundation for further segmentation and analysis in clinical applications.

Modified ResNet-34 exhibited the highest TP coverage (0.885), indicating a strong ability to localize the target region accurately. This model also achieved the highest classification accuracy (0.979); however, this improvement increased inference time (2.427s), which is a critical parameter for real-time applications.

Modified ResNet-18 also achieved high classification accuracy (0.976) while maintaining a relatively low inference time (1.461s). These results indicate that Modified ResNet-18 would be ideal for balancing accuracy and speed. However, it should be noted that TP coverage (0.815) was lower than both Modified ResNet-34 and the original ResNet-18. Considering this, it would be possible to conclude that it is a strong classifier with relatively lower localization precision.

The DenseNet model had the longest inference time (4.055s on average), making it less suitable for real-time applications. Besides being computationally expensive, its true positive coverage and classification accuracy were lower than the ResNet-based models. This shows that DenseNet was slower and less effective in detecting the seminal vesicle region. These results highlight the need for lighter and more efficient ResNet architectures for better performance.

These findings highlight a fundamental trade-off between localization performance and computational efficiency. While Modified ResNet-34 proves to be the most effective model for achieving high TP coverage, its higher inference time may limit its applicability in real-time settings. On the other hand, Original ResNet-18 remains the fastest model while maintaining competitive accuracy and TP coverage, making it a reliable choice for real-time applications. Ultimately, ResNet-based modified architectures emerge as a superior alternative to DenseNet for balancing segmentation accuracy and inference speed in seminal vesicle localization.

This study used a limited number of training samples, and no data augmentation was applied. The primary aim was to evaluate the model’s baseline capacity by observing how well it can learn from a small, unaugmented dataset. Although the sample size was limited, the dataset was curated with high anatomical precision by an expert radiologist, minimizing label noise and enhancing training effectiveness. Moreover, since a modified model was proposed, avoiding augmentation helped ensure a fair comparison with baseline architectures by eliminating external variability. It is also important to note that the seminal vesicle is a small and anatomically sensitive structure, and transformations such as rotation or flipping could have introduced spatial distortion. Despite these constraints, the model achieved strong performance. In future studies, augmentation may be explored to further enhance generalization.

Another limitation of this study is that only coronal T2-weighted MRI images were utilized. However, this was a deliberate methodological choice aimed at simplifying the problem as much as possible. Using single-axis slices, we sought to achieve the highest possible accuracy with the lowest-dimensional data, minimizing computational complexity while maintaining diagnostic effectiveness. In clinical practice, assessing seminal vesicle invasion often involves a combination of axial, sagittal, and coronal planes. While including multi-planar imaging data could further enhance structural variation analysis and overall detection accuracy, the success of our approach using only coronal slices demonstrates its feasibility as a lightweight and efficient solution. Future research could explore the integration of multi-view imaging or 3D CNN architectures to extend the current methodology while maintaining its efficiency.

The model’s reliability in detecting seminal vesicles was validated and empirically confirmed through illustrated heatmaps, which consistently focused on the expected anatomical regions. Moreover, numerical results, such as high TP coverage values, highlight the model’s ability to precisely delineate the region of interest and its robustness.

F1 score and other pixel-wise metrics such as Dice or IoU were not included in this study because the primary objective was not semantic segmentation, but rather region-level classification using a sliding-window approach. Since each 50x50 window was classified independently, performance was evaluated based on accuracy and TP coverage at both the window and image levels. Similarly, U-Net was not included in the comparison as its goal and output structure differ significantly from our patch-based method. Instead, we compared our model with ResNet and DenseNet classifiers to maintain architectural and methodological consistency. In future studies involving end-to-end segmentation, we plan to incorporate pixel-level metrics such as Dice coefficient and F1 score for a more detailed evaluation.

Identifying the region with high sensitivity is essential, as it provides a solid foundation for further delineation of anatomical boundaries. Once accurately localized, subsequent segmentation methods can refine the detection of pathological abnormalities with greater precision. Accordingly, the presented results indicate that the proposed approach could be a critical first step for accurate segmentation in abnormality detection.

Similar to prior studies utilizing deep learning for prostate cancer classification²⁶, our findings emphasize the role of AI in MRI-based diagnostics. While classification models aid clinical decision-making, precise localization is fundamental for accurate segmentation and lesion detection. This study further underscores the importance of automated localization as a crucial step toward advancing MRI-based diagnostic models.

Conclusion and future directions

This study presents a modified ResNet-based approach for automated localization of the seminal vesicle region in T2-weighted MRI scans. The applied architectural modifications improved the model’s learning capacity and stability, enhancing localization accuracy with minimal computational overhead. These findings suggest that the proposed approach is a practical and efficient solution for medical image analysis. The results indicate that the accurate localization mechanism can serve as a crucial first step toward refining segmentation and abnormality detection in MRI-based diagnostics. The proposed method reduces manual effort and interobserver variability, making it a valuable tool for clinical workflows. Future research may focus on multimodal imaging strategies that integrate axial, sagittal, and coronal planes to improve functionality. The following stages of this project could involve abnormality detection within the localized regions, for increasing its clinical value. At that stage, the integration of transformer-based or hybrid architectures with ResNet backbones may further enhance accuracy. Ultimately, this framework could be expanded toward complete segmentation of pathological involvement, such as seminal vesicle tumor invasion, with the primary goal of providing direct clinical utility in treatment planning and prognosis.

Author contributions

E.H. curated the data, Ç.C. implemented the software, E.H. and Ç.C. analyzed the results, All authors reviewed the manuscript.

Funding

This research received no external funding.

Data availability

The dataset analysed during the current study are available from the corresponding author on reasonable request. The data are not publicly available due to privacy and ethical restrictions.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

All experimental protocols were approved by the Institutional Review Board of Izmir City Hospital (approval number: 2025/49, Date: 05/02/2025).

Informed consent

Informed consent was obtained from all subjects involved in the study.