Clinical Problem Accurately measuring multiple lung cancer lesions is critical for treatment response assessment. Manual measurement by radiologists is time-consuming, subject to inter-reader variability, and particularly challenging when patients have many lesions across multiple CT slices.
Study Scale This study developed and validated a deep learning segmentation pipeline using 1,081 chest CT scans containing a total of 5,322 annotated lesions collected from multiple medical centers, making it one of the largest multi-lesion lung cancer segmentation datasets reported.
Pipeline Architecture The LLSB_CFPR pipeline combined an nnU-Net (no-new-Net, an automated architecture selection framework) backbone with a Cascade False Positive Reduction (CFPR) module using two specialized classifiers to eliminate spurious detections.
Performance On the test set, the pipeline achieved a Dice Similarity Coefficient (DSC) of 76% and sensitivity of 85%. On external validation, DSC was 73% and sensitivity remained at 85%, demonstrating strong generalizability across different institutions and scanners.
nnU-Net Framework The nnU-Net (no-new-Net) automatically configures its architecture, preprocessing steps, and training procedures based on the specific dataset properties. This self-adapting approach removes the need for manual hyperparameter tuning and was selected as the segmentation backbone.
Thoracic Bounding Box Preprocessing Before segmentation, CT volumes were cropped to a thoracic bounding box, reducing computational load and focusing the network on the relevant anatomical region. This preprocessing step improved both training efficiency and segmentation accuracy.
Cascade False Positive Reduction The CFPR module addressed the high false positive rate inherent in initial segmentation. It used two sequential classifiers: XPC (cross-patch classifier) examined local patch context, and LVC (lesion volume classifier) used volumetric features to reject false positives that passed the first stage.
RECIST 1.1 Compatibility The pipeline was specifically designed to detect measurable lesions per RECIST (Response Evaluation Criteria in Solid Tumors) 1.1 guidelines, which require lesion longest diameter of at least 10 mm. This clinical standard guided training and evaluation.
Dataset Scale 1,081 CT scans from multiple medical centers were included, with each scan containing 1-20 lung lesions. The total of 5,322 annotated lesions represents a scale of annotation effort rarely seen in academic deep learning studies.
Annotation Protocol Each lesion was manually segmented by trained radiologists using 3D volumetric segmentation tools. A senior radiologist reviewed all annotations for quality control. Inter-annotator consistency was measured using ICC to ensure reliable ground truth.
Center Diversity Including scans from multiple CT manufacturers, reconstruction kernels, and patient populations mimics real-world clinical deployment conditions. Models trained and validated on such diverse data are much more likely to generalize than single-center studies.
Lesion Characteristics The dataset included lesions spanning a wide range of sizes, densities (solid, subsolid, ground-glass), and locations (central, peripheral, juxta-pleural), covering the full spectrum of lung cancer appearances encountered in clinical practice.
Dice Similarity Coefficient DSC measures the overlap between the AI segmentation and the radiologist annotation (ground truth). A DSC of 76% on the test set and 73% on external validation indicates high spatial correspondence, with the difference between test and external sets being small and acceptable.
Sensitivity Analysis Sensitivity of 85% at both test and external validation means the pipeline correctly detected 85% of ground truth lesions. The remaining 15% were mainly very small lesions (below 10 mm, outside RECIST criteria) and lesions with ground-glass morphology.
False Positive Rate The CFPR module substantially reduced false positives compared to the nnU-Net baseline alone. Without CFPR, there were approximately 2-3 false positives per scan on average; with CFPR, this was reduced to below 0.5 per scan.
Lesion Size Stratification Performance was best for solid lesions between 10 and 30 mm in diameter. Ground-glass and partly solid lesions showed lower DSC, consistent with their inherently ill-defined boundaries that challenge even expert human readers.
RECIST Measurement Automation The most immediate application is automating RECIST 1.1 measurements in clinical trials and routine practice. Currently, radiologists manually measure the longest diameter of target lesions on every follow-up scan -- a process that is time-consuming and variable.
Consistency Across Time Points AI segmentation provides consistent lesion delineation across serial CT scans, removing the inter-reader variability that can make tumor response assessments unreliable. This is particularly valuable in multi-reader clinical trial settings.
Multi-Lesion Management Patients with oligometastatic or stage IV lung cancer may have numerous measurable lesions requiring tracking. The pipeline efficiently handles multi-lesion cases that are most burdensome for human readers.
Integration with Treatment Planning Precise volumetric segmentation from the pipeline could feed directly into radiation therapy planning systems, enabling accurate GTV delineation for patients receiving ablative radiation treatment.
Ground-Glass Lesion Limitations The pipeline underperformed on ground-glass opacity lesions, which have inherently indistinct boundaries. Specialized preprocessing or separate models trained specifically on GGO lesions may be needed.
Annotation Variability Despite quality control, inter-annotator variability in manual segmentation creates noise in the ground truth that may limit the DSC ceiling the model can achieve. Consensus annotation protocols or automated annotation assistance could improve ground truth quality.
Non-Cancer Lesions The pipeline was trained on confirmed cancer cases and may generate false positives for benign lesions (scars, granulomas, infectious nodules) in real-world clinical deployment where not all nodules are malignant.
Longitudinal Integration Future versions of the pipeline should incorporate temporal information from prior CT scans, enabling the AI to track lesion growth or shrinkage over time -- a feature essential for real-world treatment response monitoring.