How the Model Works

–

nodule-level AUC ·

–

Accuracy at a 50% threshold

–

Patient-level AUC (most suspicious nodule)

–

Sensitivity: malignant nodules caught

–

Specificity: benign nodules correctly cleared

–

Neural networks in the ensemble

ROC curve

How many malignant nodules are caught (up) for each rate of false alarms (right). A perfect model hugs the top-left corner.

Nodule levelPatient level

Show as a table

Is it overfitting?

Loading…

Show as a table

Comparing the models

A 3D ResNet trained from scratch, a 2.5D multi-view CNN built on an ImageNet-pretrained ResNet18, and their ensemble (averaged probabilities). All are measured on the same held-out patients.

Tested against real diagnoses

For LIDC patients, the diagnosis was confirmed: (methods: ). This checks whether the model predicts actual cancer, not just radiologists' opinions. Each patient is scored by their most suspicious nodule.

–

AI model AUC ·

–

Radiologists' ratings AUC ·

Predicting the true diagnosis is much harder than predicting radiologists' ratings. The model learned from those ratings, and many of these patients had cancer that had spread from elsewhere. The radiologists also saw the whole scan and its context. With only a small number of patients the confidence intervals are wide. This is the most honest measure on this page.

Results per fold

Each fold's model is tested on a different fifth of the patients; the spread shows how stable the results are.

Architecture

Two different networks, one ensemble

3D ResNet (8.3 M parameters, trained from scratch) reads the nodule as a true 3D volume, shown below. 2.5D multi-view CNN (11.2 M parameters) cuts 9 planes through the nodule (3 orthogonal, 6 diagonal), passes each through an ImageNet-pretrained ResNet18, and pools what it sees. Because they look at nodules differently, their mistakes differ, and averaging them helps.

Input48×48×48 mm cube, 1 mm voxels

Stem3D conv, 32 filters

Stage 1–2Residual blocks, 32→64 filters

Stage 3–4Residual blocks, 128→256 filters

PoolingGlobal average + dropout

OutputProbability of "malignant"

Data

Dataset: LIDC-IDRI (The Cancer Imaging Archive), 1,018 CT scans
Nodules used: – nodules from – patients, – malignant
Labels: Up to 4 radiologists rated each nodule 1–5. Mean ≤ 2.5 = benign, ≥ 3.5 = malignant; unclear nodules (in between) left out.
Pre-processing: Resampled to 1 mm, clipped to −1000…400 Hounsfield units, 64 mm cube around each nodule.

Training

Validation: 5-fold cross-validation, grouped by patient so no patient is in both training and testing
Augmentation: Random 3D crops, flips and 90° rotations
Optimisation: AdamW, cosine learning-rate schedule, 60 epochs, class-balanced loss, mixed precision
Hardware: One laptop GPU (RTX 4050, 6 GB): about 20 minutes per fold
Explainability: Grad-CAM heatmaps averaged over the 3D models
Detection: MONAI's RetinaNet trained on LUNA16 finds nodule candidates across the whole scan

Limitations

What the numbers don't tell you

Labels are opinions, not biopsies

The model learned to predict what radiologists would think about a nodule, not confirmed diagnoses. Its ceiling is their agreement with each other.

One dataset, one era

LIDC-IDRI scans come from a few US institutions in the 2000s. Scanners, protocols and populations elsewhere differ, and performance may drop there.

Detection can miss things

The automatic detector finds most nodules but not all, and sometimes flags other structures. A missed nodule gets no score, and anything you click still gets one.

No history, no growth

Doctors use growth over time, symptoms, smoking history and other tests. The model sees one 48 mm cube from one scan.

Not clinically validated. It must not be used for diagnosis or treatment decisions. Always consult a doctor.