A HYBRID RESNET-U-NET ARCHITECTURE FOR DETECTION OF LYMPHANGIOLEIOMYOMATOSIS (LAM) IN COMPUTED TOMOGRAPHY LUNG SCANS

Lymphangioleiomyomatosis (LAM) is a rare and progressive lung disease marked by the abnormal proliferation of smooth muscle-like cells and the formation of numerous thin-walled cysts within the lung tissue. Predominantly affecting women of reproductive age, LAM presents significant diagnostic challenges due to its subtle and diffuse appearance on computed tomography (CT) scans. Early and accurate identification is vital for effective disease management. A novel hybrid deep learning architecture is presented here that combines a Residual Neural Network (ResNet) with a U-Net segmentation model to automate the detection of LAM from CT lung images. In the proposed framework, ResNet functions as the encoder to extract deep semantic features, while U-Net’s decoder reconstructs precise spatial representations through skip connections. This Framework increases the effectiveness of segmenting and detecting cystic abnormalities in LAM cases. The resulting segmentation maps offer detailed anatomical insights, which can be leveraged to extract morphological features for auxiliary classification, further distinguishing LAM from other cystic lung pathologies. Extensive experiments were conducted using a curated and annotated dataset of CT lung scans. The hybrid model demonstrated superior segmentation performance, achieving high Dice coefficients and improved classification accuracy. Compared to traditional U-Net and standalone CNN approaches, the ResNet-U-Net architecture yielded significant gains in sensitivity, specificity, and overall robustness. This study highlights the potential of integrating residual learning with encoder–decoder-based segmentation for rare disease detection and illustrates the promise of AI-powered tools in enhancing diagnostic accuracy within radiological workflows.