NEW CUTTING-edge deep-learning model using lung CT imaging has shown significant potential for accurately identifying and segmenting lung tumours, according to researchers based in the USA. The findings could have transformative implications for lung cancer treatment.
“This study represents an important step toward automating lung tumour identification and segmentation,” said the research team in a statement released by the Radiological Society of North America (RSNA). “This approach could impact areas like automated treatment planning, tumour burden quantification, and treatment response assessment.”
Manually delineating lung tumours on CT scans is a critical but labour-intensive process subject to variability among physicians. To address this, the group developed a 3D U-Net-based ensemble deep-learning model, designed to identify and segment tumours efficiently and consistently.
The model was trained using data from 1,504 CT scans and clinical tumour segmentations of 1,295 patients treated with radiotherapy for primary or metastatic lung tumours. Researchers evaluated the model on sensitivity, specificity, false positive rate, and Dice similarity coefficient (DSC), comparing its performance with that of physicians.
The study reported promising results:
- The model demonstrated 92% sensitivity and 82% specificity in detecting lung tumours.
- On a subset of 100 scans, the model achieved a median Dice similarity coefficient of 0.77, comparable to the interphysician DSC of 0.8.
- Segmentation was faster with the model, averaging 76.6 seconds per scan compared to 166.1–187.7 seconds for physicians.
“By capturing rich interslice information, our 3D model is theoretically capable of identifying smaller lesions that 2D models may overlook,” the authors noted.
The team hopes their scalable data extraction pipeline can inspire other radiology departments to create their own medical image segmentation datasets, improving efficiency and accuracy across institutions.
These advancements could play a pivotal role in the future of cancer treatment, providing faster and more reliable data for clinical decision-making.
Reference
Kashyap M et al. Automated deep learning-based detection and segmentation of lung tumors at CT. Radiology. 2025;314(1):e233029.
