Accurate kidney tumor semantic segmentation.

Abstract

Timely and accurate detection of kidney tumors is crucial for improving clinical outcomes, as early diagnosis significantly reduces the risk of progression to advanced stages. Kidney tumors, particularly malignant ones, can be life-threatening if not identified early. Developing a robust, efficient, and automated framework for detecting and classifying kidney tumors from CT images is imperative to enable early intervention and improve survival rates. This study addresses critical challenges in kidney tumor detection, such as variability in tumor characteristics, limited availability of medical experts, and the timeintensive nature of the manual diagnosis. It explores advanced deep learning models to automate and enhance the accuracy and speed of tumor segmentation, offering a solution to bridge the gap in current diagnostic approaches. The research evaluates the performance of multiple architectures, including the EfficientNet family, SE-ResNet models with Feature Pyramid Network (FPN), and DenseNet integrated with LinkNet. These models were assessed based on key metrics like Intersection over Union (IoU), Dice similarity coefficients, and computational efficiency. EfficientNet-B4 demonstrated exceptional tumor segmentation capabilities, achieving a Tumor Dice score of 0.984 and Mean Dice of 0.990, while EfficientNet-B7 excelled in kidney segmentation with IoU and Dice scores of 0.977 and 0.988, respectively. SE-ResNet50 emerged as the top-performing model in overall segmentation, achieving a Tumor Dice score of 0.992 and a Mean IoU of 0.988. DenseNet121 with LinkNet offered a balance of precision and computational efficiency, achieving Tumor IoU and Dice scores of 0.954 and 0.976, respectively. This study highlights the potential of integrating scalable deep learning architectures into clinical workflows. EfficientNet models stand out for their resource-constrained adaptability, while SE-ResNet50 demonstrates superior precision, particularly in contexts requiring detailed segmentation. DenseNet121 combines stability and efficiency, offering a versatile solution for varying clinical requirements. The findings underscore the importance of leveraging advanced architectures to enhance kidney cancer diagnostics, paving the way for further optimization and broader real-world deployment. This contribution establishes benchmarks in automated tumor detection and segmentation, providing a foundation for future research aimed at improving diagnostic accuracy and patient outcomes in resourceconstrained and high-performance clinical environments.