Linear Fusion of Serial Screening CT for Lung Nodule Malignancy Prediction

Lung cancer screening is longitudinal by design: participants return every year, and radiologists look at how a nodule has changed since the last visit before deciding what to do next. Most deep learning models for nodule malignancy ignore this and are trained and evaluated on a single computed tomography (CT) scan. The evidence that prior imaging helps comes mostly from small studies. Paez et al. [1] fitted a joint model over serial scores from a commercial classifier in 80 patients (32 malignant), and found that the longitudinal scores discriminated better than the most recent score alone. This study asks how much a prior scan actually adds at screening scale. This study used the LUNA25 public training and development set [2]: 6,163 annotated nodules from 4,069 low-dose CT scans in 2,120 National Lung Screening Trial participants. Malignant nodules were annotated by an experienced radiologist and confirmed histopathologically; benign nodules were annotated by trained medical students and verified by at least two years of imaging stability. Grouping annotations by lesion leaves 3,762 nodules, and 1,603 of these (149 malignant) were imaged at two or three annual rounds. Malignancy prevalence is flat across sequence lengths, so having follow-up imaging is not itself tied to outcome. Slice thickness in this multicentre cohort ranges from 0.5 to 3.2 mm, so a fixed 48 mm physical cube was cropped around each lesion centre and resampled to 64³ isotropic voxels. A nodule of a given physical size then occupies the same voxel count regardless of scanner. A 3D convolutional network was trained with patient-grouped, stratified five-fold cross-validation, with no patient contributing to more than one fold, and it reached a pooled AUC of 0.8807 (per-fold mean 0.885, standard deviation 0.015). This study then asks what this single-timepoint network already achieves when handed a whole sequence. Averaging its out-of-fold predictions across timepoints gives 0.8796 and taking the maximum gives 0.9108, but simply using the most recent scan gives 0.9251, even though the network never sees the order of the scans. Recency alone is therefore a stronger baseline than the aggregations longitudinal models are usually compared against, so it is adopted as the reference. A logistic regression on the first and last scores reaches 0.9433, an improvement of +0.0182 over that reference (95% CI +0.0079 to +0.0287, p = 0.0012, paired bootstrap over 5,000 resamples). Later scans are individually more informative (AUC 0.8288, 0.8696 and 0.9247 for the first, second and third of three annual rounds), consistent with lesions becoming easier to classify nearer diagnosis. The advantage of combining scans is not explained by this: the first scan is the weakest single predictor, yet adding it to the most recent one improves discrimination, and the gain is present in both two-scan (+0.0253) and three-scan (+0.0089) sequences. Adding the mean, maximum, standard deviation, slope and elapsed interval changes nothing (0.9420). Logistic regression beat random forest, gradient boosting and a multilayer perceptron in all four feature sets tested, and age and sex, the only patient covariates released with LUNA25, are uninformative on their own (AUC 0.5321). A prior scan therefore gives a real and statistically significant but modest advantage, and two scans with a linear model capture all of it. The main limitations are that only 149 malignant sequences are available, that single-timepoint predictions come from best-validation-epoch checkpoints, and that no external validation is reported.