Two tasks that should work together. Lung cancer screening with LDCT requires two successive decisions: first, detecting that a nodule is present and locating it precisely; second, estimating whether that nodule is likely malignant. Most AI systems treat these as separate problems, which introduces redundancy and can create inconsistencies between where a nodule is located and what malignancy probability it receives.
Seg-CADe-CADx integrates both tasks. This study proposes a unified two-stage deep learning framework that processes CT scans end-to-end: Stage 1 (Seg-CADe) detects and localizes nodules using segmentation guidance; Stage 2 (CADx) classifies the malignancy risk of detected candidates using a hybrid architecture combining CNN and Transformer models. An additional calibration step ensures the output probability scores are clinically trustworthy.
The clinical problem this addresses. Radiologists face thousands of CT slices per scan, where nodules can be small, subsolid, or vessel-attached and easily missed. High false-positive rates from AI detectors create unnecessary follow-up workload. The goal is a system with high sensitivity for small nodules, low false positives, and accurately calibrated malignancy probabilities - all in a single integrated pipeline.
Validated on standard public benchmarks. The framework was evaluated on LUNA16 (888 CT scans, used for detection benchmarking) and LIDC-IDRI (1,018 CT scans, used for malignancy classification), as well as on held-out NLST data for external validation under realistic clinical domain shift.
Segmentation as a detection guide. Rather than directly predicting nodule locations, Stage 1 first trains a residual 3D U-Net to produce a voxel-by-voxel probability map of where nodules are likely to exist. This map then guides a downstream detection head, ensuring that candidate proposals come from anatomically plausible lung tissue and reducing false positives from vessels and artifacts outside the lung.
Handling the class imbalance problem. Nodule voxels represent a tiny fraction of the total CT volume - a severe class imbalance that can cause detectors to ignore small nodules entirely. Three strategies address this: a 2:1 ratio of negative-to-positive patch sampling during training, foreground voxel weighting three times higher than background in the loss function, and a joint Dice-plus-Focal loss that emphasizes boundary learning.
The 2.5D refinement head. After the 3D U-Net generates candidate nodule locations, a lightweight 2.5D refinement head (processing stacks of adjacent 2D slices) refines the localization by predicting precise nodule center coordinates, radii, and classification scores. False positives are suppressed through anatomical masking restricted to the lung region and hard-negative mining to penalize recurring false detections.
Segmentation quality and detection performance. The 3D U-Net achieved a mean Dice score of 0.742 on the test set - modest but sufficient to guide the downstream detector. Importantly, the segmentation does not need to be perfect to function as an effective detection guide; it only needs to reliably flag anatomically plausible nodule regions. The final detector achieved CPM of 0.944 and 96.4% sensitivity at 1 false positive per scan.
Two complementary neural network streams. Stage 2 processes each detected nodule candidate through two parallel architectures. A 3D DenseNet backbone captures fine-grained local texture details - the subtle differences in density patterns, edge characteristics, and internal structure that distinguish benign from malignant tissue. Simultaneously, a 3D Swin Transformer models long-range contextual relationships across the entire nodule volume, capturing global patterns that local convolutions miss.
Feature fusion and multi-task prediction. The feature embeddings from both streams are concatenated and fed through a multilayer perceptron to produce a unified representation. The classifier then outputs both a malignancy probability and auxiliary predictions for radiological attributes such as spiculation (spike-like margins), lobulation (scalloped edges), and subtlety - all of which contribute to the interpretability of the prediction.
Radiomics features added for interpretability. A compact set of handcrafted radiomic descriptors - nodule sphericity, compactness, surface-to-volume ratio, and intensity statistics - is extracted from the segmented nodule and fused with the deep learning embeddings. This late-fusion approach adds interpretable, clinician-recognizable features to the model's decision without requiring them to replace learned deep features.
Probability calibration for clinical trust. Raw deep learning outputs are often poorly calibrated - a model may output a 90% malignancy probability for nodules that are actually malignant only 70% of the time. Temperature scaling is applied to align predicted probabilities with true malignancy likelihoods. Monte Carlo dropout layers further provide uncertainty estimates alongside each prediction, supporting appropriate clinical confidence.
Best-in-class detection performance. On the LUNA16 benchmark, Seg-CADe-CADx achieved a Competition Performance Metric (CPM) of 0.944 and 96.4% sensitivity at 1 false positive per scan - outperforming all compared methods including SCPM-Net (0.892), multi-scale attention detectors (0.927), and Faster R-CNN + OHEM (0.901). The improvement was statistically significant (p less than 0.05).
Strong performance specifically for small nodules. Size-stratified analysis shows sensitivity of 91.0% for nodules 6 mm or smaller - the most clinically important and most challenging size range for early cancer detection. Sensitivity increased to 95.5% for 6-10 mm nodules and 98.9% for nodules larger than 10 mm, confirming that the segmentation-guided approach specifically addresses the small nodule detection gap.
Malignancy classification: near-perfect AUC. On LIDC-IDRI, the hybrid classifier achieved ROC-AUC of 0.988 and PR-AUC of 0.947. At the Youden optimal threshold, sensitivity was 96.6% and specificity 92.9%. At a clinically important operating point of 95% sensitivity, specificity remained at 93.8% - meaning that even while catching nearly all malignant nodules, the system correctly identifies over 93% of benign ones as benign.
Outperforms all recent comparison methods. The classifier surpassed NoduleX (AUC 0.990, Acc 92.0%), 3D-MCN Capsule (AUC 0.964, Acc 93.1%), MTST-Swin (AUC 0.982, Acc 93.7%), and GC-WIR Attention (AUC 0.985, Acc 94.3%), achieving 96.9% accuracy, 95.8% sensitivity, and 97.8% specificity. The improvement over all baselines was statistically significant (p less than 0.01 for AUC).
What calibration means in clinical practice. A well-calibrated model outputs probabilities that match reality: if it says a nodule has a 30% malignancy probability, roughly 30% of such nodules should actually be malignant. Miscalibrated models create clinical problems - an overconfident model triggers unnecessary biopsies; an underconfident one may delay urgent referrals.
Calibration performance. After temperature scaling, Seg-CADe-CADx achieved an Expected Calibration Error (ECE) of 0.209 and a Brier score of 0.083 - a 34% improvement over the uncalibrated version (ECE 0.317, Brier 0.112). CNN-only and Swin-only classifiers had ECE of 0.284 and 0.261 respectively, showing that hybrid feature integration benefits both discrimination and calibration reliability.
Decision curve analysis confirms clinical benefit. Decision curve analysis (DCA) quantifies net clinical benefit across the range of threshold probabilities used in clinical decision-making. The calibrated Seg-CADe-CADx model consistently outperformed uncalibrated predictions, as well as the 'treat all' and 'treat none' strategies, across the 10-40% threshold range most relevant to follow-up scheduling and biopsy referral decisions.
Fewer unnecessary biopsies with calibration. By aligning predicted probabilities with true malignancy risk, calibration reduces the proportion of benign nodules incorrectly assigned to high-risk categories. This directly translates to fewer unnecessary biopsies, reduced patient anxiety, lower healthcare costs, and reduced radiation exposure from avoidable follow-up imaging.
CNN and Transformer provide complementary strengths. When tested individually on LIDC-IDRI, DenseNet alone achieved AUC 0.972 and Swin Transformer alone achieved AUC 0.981. Their hybrid combination reached AUC 0.993 - better than either alone, confirming that local texture (CNN) and global context (Transformer) capture genuinely different discriminative signals in nodule images.
Radiomics adds interpretability at a small accuracy cost. Adding radiomics features to the hybrid model slightly reduced AUC from 0.993 to 0.988 but improved calibration stability. This trade-off is clinically desirable: the marginally lower discrimination accuracy is offset by increased interpretability, helping clinicians understand why the model assigned a particular malignancy score.
Segmentation guidance is essential for detection. The ablation study showed statistically significant performance drops when segmentation guidance was removed (p less than 0.01 for both AUC and CPM). Without segmentation-guided proposals, the detector misses more small nodules and generates more false positives, confirming that the anatomical context provided by the U-Net is a critical component of the pipeline.
Model interpretability through Grad-CAM and SHAP. Attention heatmaps from Grad-CAM confirm that the hybrid classifier focuses on nodule margins and internal texture for malignant predictions, and on smooth boundaries for benign ones - matching radiologist decision-making patterns. SHAP attribution maps corroborate these findings, showing positive contributions from irregular edges and asymmetric density in malignant cases.
Tested on NLST without fine-tuning. External validation on the National Lung Screening Trial (NLST) cohort - acquired at multiple clinical centers with different scanner vendors and protocols - showed CPM 0.903 and ROC-AUC 0.951 without any additional retraining. This represents a modest but expected performance drop from domain shift, while retaining clinically acceptable performance across a genuinely new patient population.
Practical deployment on standard hardware. The full pipeline processes one CT scan in approximately 4 seconds with peak GPU memory of 9.8 GB - well within the constraints of standard clinical workstation GPUs. A lighter 'Lite' variant achieves a 34% runtime reduction with only 1.5% AUC loss, making the system viable for deployment on lower-powered edge devices or portable diagnostic units.
Common failure modes identified. Error analysis reveals that false positives primarily arise from vessel-like structures and imaging artifacts that resemble nodules in appearance. False negatives occur mainly with very small or very low-contrast nodules - the same cases that are challenging for human radiologists. These failure patterns provide clear targets for future algorithmic improvements.
Integration into existing clinical workflows. Processing at 4 seconds per scan is compatible with typical PACS-integrated CAD systems (which generally allow 10-15 seconds). The calibrated malignancy probabilities, combined with visual explanations from Grad-CAM and SHAP, provide the radiologist with both a risk score and a visual indication of which nodule features drove the score - supporting rather than replacing clinical judgment.