Medical and radiation science analysis using a probability U-Net with a GRU model

Accurate and interpretable segmentation of CT images of the lungs is critical for early diagnosis and treatment planning in pulmonary diseases. The current deep learning (DL) models tend to lack uncertainty quantification, lack consistency over time, and do not extrapolate into ambiguous areas. To solve these problems, this study suggests a hybrid DL model that combines U-Net with Gated Recurrent Units (GRUs) and a probabilistic latent space to segment based on uncertainty. The encoder learns both spatial and temporal features using GRU-enhanced convolutional blocks, and the probabilistic module of the bottleneck addresses aleatory and epistemic uncertainty via Monte Carlo sampling. This joint temporal modeling and probabilistic representation allows the system to perform temporally consistent segmentation with only one lightweight architecture. The experiments on the IQ-OTH/NCCD lung CT dataset (1295 annotated images) demonstrate that the proposed model achieves 98.9 % accuracy and a 91.5 % Dice coefficient, which is 7–15 % higher than state-of-the-art models, including ResNet18 with MC Dropout and SkinSAM. The model also achieves an Expected Calibration Error of just 2.8 % which justifies its validity in the clinic. Uncertainty maps can also aid decision-making regarding tumor boundaries. Compared with previous applications such as Probabilistic U-Net and GRU-Net, which operate independently, the suggested U-Net-GRU-Probabilistic framework offers an integrated, confident, and understandable method for segmenting lung CT. Demonstrates potential for real-time diagnostic deployment.