Abstract
Gastrointestinal (GI) cancer is a fatal malignancy that affects the organs of the GI tract. The rising prevalence of GI cancer has recently influenced the health of millions of people. To treat GI cancer, radiation oncologists must carefully focus X-rays on tumors while avoiding other unaffected organs in the GI tract. This research proposes a novel approach to segment healthy organs within the GI tract from magnetic resonance imaging (MRI) scans using a multi-level attention DeepLab V3 + model. The proposed model aims to enhance segmentation performance by incorporating state-of-the-art approaches, such as atrous convolutions and EfficientNet B0 as an encoder, by leveraging hierarchical information present in the data. Here, the attention mechanism is applied at multiple levels of features, i.e., low, medium, and high, to capture and leverage hierarchical information present in the data. At the same time, EfficientNet B0 extracts deep and meaningful features from input images, providing a robust representation of GI tract structures. Hierarchical feature fusion combines local and global contextual information, resulting in more accurate segmentation with fine-grained details. The model is implemented using the UW-Madison dataset, comprising MRI scans from 85 patients with gastrointestinal cancer. To optimize the model, it has been simulated with different parameters, including optimizers, the number of epochs, and cross-validation folds. The model has achieved performance metrics such as a model loss of 0.0044, a dice coefficient of 0.9378, and an Intersection over Union (IoU) of 0.921.