Abstract
Cancer is an abnormal growth with potential to invade locally and metastasizeto distant organs. Accurate auto-segmentation of the tumor and surroundingnormal tissues is required for radiotherapy treatment plan optimization. RecentAI-based segmentation models are generally trained on large public datasets,which lack the heterogeneity of local patient populations. While these studiesadvance AI-based medical image segmentation, research on local datasets isnecessary to develop and integrate AI tumor segmentation models directly intohospital software for efficient and accurate oncology treatment planning andexecution. This study enhances tumor segmentation using computationallyefficient hybrid UNet-Transformer models on magnetic resonance imaging (MRI)datasets acquired from a local hospital under strict privacy protection. Wedeveloped a robust data pipeline for seamless DICOM extraction andpreprocessing, followed by extensive image augmentation to ensure modelgeneralization across diverse clinical settings, resulting in a total datasetof 6080 images for training. Our novel architecture integrates UNet-basedconvolutional neural networks with a transformer bottleneck and complementaryattention modules, including efficient attention, Squeeze-and-Excitation (SE)blocks, Convolutional Block Attention Module (CBAM), and ResNeXt blocks. Toaccelerate convergence and reduce computational demands, we used a maximumbatch size of 8 and initialized the encoder with pretrained ImageNet weights,training the model on dual NVIDIA T4 GPUs via checkpointing to overcomeKaggle's runtime limits. Quantitative evaluation on the local MRI datasetyielded a Dice similarity coefficient of 0.764 and an Intersection over Union(IoU) of 0.736, demonstrating competitive performance despite limited data andunderscoring the importance of site-specific model development for clinicaldeployment.