Trans RCED-UNet3+: A Hybrid CNN-Transformer Model for Precise Lung Nodule Segmentation

Sadaf Raza Sadaf^1*Razia Zia^2,3Irfan Ahmed Usmani⁴Nouf Almujally⁵Nada Alasbali⁶Muhammad Hanif^7*

• ¹Sir Syed University of Engineering and Technology, Karachi, Pakistan
• ²Iqra University, Karachi, Pakistan
• ³Department of Computer Science Iqra University, Karachi, Pakistan
• ⁴Salim Habib University, Karachi, Pakistan
• ⁵Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia
• ⁶King Khalid University College of Medicine, Abha, Saudi Arabia
• ⁷Orebro universitet, Örebro, Sweden

Precisely segmenting lung nodules in CT scans is essential for diagnosing lung cancer, though it is challenging due to the small size and intricate shapes of these nodules. This study presents Trans RCED-UNet3+, an enhanced version of the RCED-UNet3+ framework designed to address these challenges. The model features a transformer-based bottleneck that captures global context and long-range dependencies, along with residual connections that facilitate efficient feature flow and prevent gradient loss. To improve boundary accuracy, we employ a hybrid loss function that combines Dice loss with Binary Cross-Entropy, enhancing the clarity of nodule edges. Evaluation on the LIDC-IDRI dataset demonstrates a notable advancement, as Trans RCED-UNet3+ achieves a Dice score of 0.990, exceeding the original model's score of 0.984. These findings underscore the value of merging convolutional and transformer architectures, delivering a robust approach for precise segmentation in medical imaging. This model enhances the detection of subtle and irregular structures, enabling more accurate lung cancer diagnoses in clinical environments.