Gastrointestinal cancers include cancers of the colon, liver, stomach, and esophagus, which are the most common and deadly in the world [1]. They are a substantial source of health burden, especially among older men, and have led to high mortality rate around the globe. GLOBOCAN, the International Agency for Research on Cancer reported that cancer remains one of the most common diseases and causes of death, accounting for around 1.93 million new cases in 2020 and 900,000 deaths. The statistics unveil essential issues that should be urgently addressed through preventive measures, proper early detection methods, and better treatment protocols to mitigate the global burden of cancer [2].

The treatment of GI cancers is generally wide-ranging and based on the type of cancer. For example, patients who have colon cancer often experience surgical intervention, chemotherapy, or radiation therapy [3, 4]. To be more specific, among the three, it tends to be central with the position of being more dominant because it utilizes high-energy X-rays to eliminate cancer cells [4]. Radiation therapy poses a significant challenge in the GI tract. It enables radiation oncologists to reach nearer the cancer cells without affecting the rest of the healthy tissues [5, 6]. Determining tumor size and location helps to optimize treatment plans by giving a higher radiation dose to cancerous tissues, thus providing an effective and targeted approach. Segmentation further allows for easy follow-up of the treatment response, where clinicians can analyze changes in the affected organs’ size and shape during therapy [7, 8].

There has been a revolution in clinical practice over the past decade with the appearance of deep learning as a drastically transforming tool, mainly in the diagnostic space of medical imaging [9, 10]. Techniques like image classification, object recognition, and segmentation have dramatically improved disease diagnosis and treatment planning, thus increasing the accuracy and personalization of patient care [11, 12]. CNN and U-Net architectures shown as deep learning models exhibited auspicious performance in segmenting small intestines, large intestines, and stomachs from MRI scans [13, 14]. This considerable dataset training facilitates the models to recognize and outline segments of disease areas, enabling clinicians to establish relevant information for early detection, treatment, and follow-up monitoring [15–17].

This study introduces a novel deep-learning model that integrates Swin Transformer Blocks and U-Net architecture for the semantic segmentation of GI structures, explicitly targeting the small intestine, large intestine, and stomach. The model leverages the strengths of U-Net, which is optimized for segmentation tasks, and Swin Transformer, which effectively captures global context and pixel relationships within images. Our model achieves highly detailed and precise segmentation by combining these two approaches. The proposed model holds potential for significant clinical applications, enhancing the ability to accurately identify anatomical structures and improving diagnostic, therapeutic, and follow-up capabilities in GI cancer management. This research effort makes significant contributions as follows:

• The paper presents a novel approach that merges the Swin Transformer and U-Net models to leverage global context learning capabilities and fine-grained spatial resolution. Incorporating this integration greatly enhances the model’s capacity to achieve precise and comprehensive semantic segmentation, specifically in accurately outlining the gastrointestinal tract in MRI data.

The following outlines the later parts of this study: Section 2 summarizes the literature work, and Section 3 addresses the input dataset. Section 4 elaborates on the proposed Integrated Swin Transformer U-Net Model, Section 5 represents the results, Section 6 offers a comparative investigation of the proposed model with state-of-the-art outcomes, and Section 7 presents the conclusion.

Many medical imaging researchers have used deep learning architectures to build segmentation and classification models for the gastrointestinal system. Ganz et al. [18] developed software based on narrow-band imaging (NBI) data to differentiate polyps autonomously. The proposed model outperforms previous algorithms for automatically segmenting 87 images. Wang et al. [19] developed a technique named “Polyp-Alert” to support endoscopists in locating polyps during colonoscopy. By monitoring the detected polyp edge(s), the method aggregates images of the same polyp(s) in one shot. Vázquez et al. [20] provided an enlarged segmentation dataset intending to create a novel robust norm for research into colonoscopy image analysis. The proposed dataset includes four relevant classifications for evaluating the endoluminal scene, each serving a different therapeutic need. Using the dataset, the authors train conventional fully convolutional networks (FCNs) to construct new baselines.

Brandao et al. [21] described a DL-based segmentation algorithm for identifying lesions in colonoscopy images. Shape-from-shading is also used to provide a more comprehensive picture of tissues. Depth is introduced as an extra input passage to the RGB data in their network models, and the resulting network performs better. The segmentation model got an IoU of 48%, producing an IoU of 56.95% on the CVC-Colon dataset. Dijkstra et al. [22] described a one-step method for detecting polyps. The approach leverages an FCNN model for segmenting polyp. They tested the proposed network on different datasets, and their outcomes were promising.

Banik et al. [23] offered a multiscale patch network for automatic polyp area segmentation. The patches are then concatenated for precise polyp area pixel label annotation. The proposed model was validated using the CVC-Clinic DB. Wang et al. [24] created a multiscale MCNet for segmentation of GI Tract endoscopic images, using global and local contexts as training guidance. One global subnetwork determines each input image’s worldwide structure. They then build two cascaded local subnetworks based on the worldwide subnetwork’s output feature maps to collect regional appearance. Three subnetworks learn feature maps concatenated for the lesion segmentation task. Galdran et al. [25] described a new approach for gastrointestinal polyp delineation using an encoder-decoder approach. In the proposed method, pre-trained encoder-decoder architecture was successively joined. Sharma et al. [26] used an encoder and a standard U-Net architecture. More sophisticated algorithms with remarkable performance in various classification contexts are available. One can encode these models to generate a distinctive U-Net design and improve output. Ye et al. [27] proposed SIA-UNet, a modified network including MRI sequence information. Extensive studies on the UWM database were conducted to evaluate the suggested model. Chou et al. [28] employed Mask R-CNN along with U-Net techniques to distinguish the GI parts. Sharma et al. [29] suggested a model that is a U-Net design built from the ground up and utilized for image segmentation.

Li et al. [30] examined and combined several 2.5D data creation strategies to make the most of the images and proposed a 2.5D feature combination approach with adjacent weighting. Their solution integrates several representation processes by deeply combining multidimensional convolutions into fundamental modules. Extensive experiments on a publically accessible GI database show that the 2.5D combination strategy outperforms the 2.5D method devoid of feature combination by 0.36% on dice and 0.12% on Jaccard. Using two methods—a UNet and a ResNet50 encoder—and a sparser UNet—Chia et al. [31] look at FiLM, a technique for leveraging pixel width and height picture data to improve UNet design. Using the variety of methods of the ensemble, Georgescu et al. [32] offered a fresh strategy for building ensembles of different medical picture segmentation architectures. Choosing the structures with the highest scores reveals that DiPE surpasses several designs and ensemble-building approaches.

The UW Madison GI tract dataset is employed in the proposed study. The University of Wisconsin has released the dataset, which is available on the Kaggle platform [33]. The collection contains 38,496 MRI scans of the GI tract for actual cancer patients.The ground truth of the dataset is in RLE format (Run Length Encoding), so the ground truth mask is created using RLE decoding. The segmentation mask is divided into three classes: small bowel, large bowel, and stomach. The size of the images in a dataset is not same for all the images, so the dataset has been resized to make all the images of same size. The input size for the images is set to 240 × 240. Table 1 displays some dataset’s sample images and corresponding ground truth masks. Figures 1A, B show two MRI images. In contrast, Figures 1C, D show the ground truth masks, with yellow representing the large bowel, green representing the small bowel, and red representing the stomach. The dataset has been separated in the ratio of 70:15:15 for train, testing, and validation, respectively.

The proposed Integrated Swin Transformer U-Net Model combines the Swin Transformer design, a breakthrough in computer vision [34], with the U-Net model [35], which is well-known for its segmentation capabilities. Combining UNet with the Swin Transformer gives the benefits from both the fine-grained spatial resolution of UNet and the high-level context information of the Swin Transformer. With its unique U-shaped topology, the combination of a contracting path for context, and an extensive method for accurate localization, the U-Net architecture captures minute details. Conversely, the Swin Transformer retains localized detail extraction rapidly and adds sliding windows (hence “Swin”) to capture broader context. The Swin Transformer is advantageous over pure transformers for vision tasks due to its hierarchical feature learning and computational efficiency. Using shifted windows to capture localized self-attention reduces the quadratic complexity of processing entire images, making it more scalable and manageable for high-resolution inputs. This “shifted window” mechanism also enables Swin Transformers to capture fine-grained details and global context as information flows between neighboring windows across layers. This combination of localized and global attention makes Swin Transformers particularly effective for image segmentation, where understanding both local structures and overall context is crucial, providing a balanced and efficient approach that pure transformers lack.

Here, employing self-attention techniques to grasp the global context and long-term relationships of images, the Swin Transformer Block and U-Net model have been constructed to combine their capabilities. Its unique feature allows it to combine international academic content with local inference. The proposed model includes three main components: encoder, bottleneck layer, and decoder. In the encoder, decoder, and bottleneck part of the U-Net model, Swin transformer blocks are used to gradually reduce the spatial dimension while increasing the complexity of the received data. Between the encoder’s downsampling and the decoder’s upsampling, the bottleneck layer refines and compresses the coding features, thus allowing us to know how much information flows through the network. The U-Net model’s decoder component improves features gathered during the encoding phase, enabling the network to record fine-grained data. A detailed description of the encoder, decoder, and bottleneck block of the proposed model is given in the following sections.

This research proposed an Integrated Swin Transformer U-Net Model to segment the gastrointestinal tract with MRI data. The model runs on the Google Colab platform using Keras and TensorFlow framework. Table 1 describes the Swin Transformer U-Net model’s training parameters proposed in GI organ segmentation task. The selected batch size is 8 to balance the two objectives so that it will not lose any performance and minimize memory use. The learning rate has been set to be 0.0001, which is small enough to ensure that convergence is stable instead of overshooting, which is what matters most for such complex architecture deep learning models like Swin Transformer U-Net. The model was trained over 70 times; thus, there were more than enough iterations to fit the dataset without overfitting. This training run took 6 h, 37 min, and 43 s, demonstrating how computer-intensive training efficient models can be on vast amounts of medical data. The following section presents this model’s results and showcases how it can help segment small bowel, large bowel, and stomach from MRI images.

Table 3 provides a comparative overview of several image segmentation approaches assessed for their effectiveness in the context of a given goal, most likely in medical imaging or computer vision, in 2022. The techniques described include transfer learning encoders, U-Net architecture, Mask RCNN, a mix of U-Net and transfer learning models, U-Net applied to 2.5D images, U-Net paired with ResNet 50, and ensemble learning. The associated Dice coefficient and IoU/Jaccard scores serve as performance measures, assessing the quality of segmentation findings. Highlights include the proposed model, which has a Dice value of 0.91 and an IoU/Jaccard of 0.84, and additional algorithms with varied segmentation accuracy. In this case, the superior performance of Swin Transformer-U-Net can be attributed to the combination of global content and spatial accuracy. Swin Transformer’s moving window effectively captures surface irregularities, enabling the model to identify minor differences between similar tissues in the colon. Furthermore, the U-Net model refines region boundaries through cross-linking, essential for accurate segmentation. This combination makes the model more efficient than previous methods by evaluating local details with global context understanding and makes it particularly suitable for complex anatomical segmentation tasks.

This study presents an integrated Swin Transformer U-Net model for segmenting intestinal lesions in MRI images, which is an essential task for developing radiology in cancer treatment. The well-designed model combines the global content learning capabilities of Swin Transformer with the detailed feature extraction capabilities of U-Net to provide optimal performance. Experimental results validated in gastrointestinal diseases at the University of Wisconsin-Madison showed that the model has high accuracy with low loss, high Dice coefficient, and IoU scores of 0.0949, 0.9190, and 0.8454, respectively. These results indicate that the proposed model can improve the accuracy of GI cancer treatment and provide radiation oncologists with a powerful tool for better treatment planning and patient care. Integrating these principles into clinical practice will lead to more efficient and effective radiation therapy, ultimately improving patient outcomes. We plan to refine the model’s architecture for future enhancements to reduce computational complexity, allowing for more efficient real-time applications. We also aim to explore further multi-modal data integration, such as combining MRI with CT scans, to improve segmentation accuracy. Beyond GI tract segmentation, this model’s framework could be adapted to other types of cancer and areas of medical imaging by fine-tuning its parameters to accommodate different tissue characteristics and imaging modalities. For instance, it could be adapted for lung or brain tumor segmentation by training on specialized datasets, enabling broader clinical applications across oncology.