henhao
Sign in
BioinformaticsbioRxiv
Preprint

DeepLung: 3D Deep Convolutional Nets for Automated Pulmonary Nodule Detection and Classification

Wentao ZhuiD, Chaochun Liu, Wei FaniD, Xiaohui S. XieiD

Abstract

In this work, we present a fully automated lung CT cancer diagnosis system, DeepLung. DeepLung contains two parts, nodule detection and classification. Considering the 3D nature of lung CT data, two 3D networks are designed for the nodule detection and classification respectively. Specifically, a 3D Faster R-CNN is designed for nodule detection with a U-net-like encoder-decoder structure to effectively learn nodule features. For nodule classification, gradient boosting machine (GBM) with 3D dual path network (DPN) features is proposed. The nodule classification subnetwork is validated on a public dataset from LIDC-IDRI, on which it achieves better performance than state-of-the-art approaches, and surpasses the average performance of four experienced doctors. For the DeepLung system, candidate nodules are detected first by the nodule detection subnetwork, and nodule diagnosis is conducted by the classification subnetwork. Extensive experimental results demonstrate the DeepLung is comparable to the experienced doctors both for the nodule-level and patient-level diagnosis on the LIDC-IDRI dataset.

PDFDOI
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
80
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (17 Sept 2017) · CC BY 4.0

On LIDC-IDRI, gradient boosting on 3D dual path network features reaches 90.44% accuracy on annotated nodules; with a 3D Faster R-CNN detector, patient-level accuracy is 81.41%, 99% of doctors' average.

The authors say lung CT work has mostly handled nodule detection and nodule classification separately; they present DeepLung as a complete automated pipeline, which they call the first, and compare it with experienced doctors. The new parts are a 3D Faster R-CNN with a U-net-like encoder-decoder that skips a separate false-positive reduction stage, and a gradient boosting classifier built on 3D dual path network features.

Key results

  1. ShownThe 3D Faster R-CNN detector comes close to the best two-stage method on LUNA16, even though it has no separate step for removing false positives.

    FROC 83.4% vs 83.9% for the best (two-stage) method, using 10-fold patient-level cross validation on 888 LUNA16 CTs; recall of 94.6% across all nodules; Fig. 5 shows 95% confidence bounds from 1,000 bootstraps. · DeepLung for Nodule Detection; Datasets; Figure legends

  2. ShownThe best malignancy classifier in the comparison was gradient boosting on 3D dual path network features (2,560 dimensions) combined with nodule size and raw 16×16×16 nodule pixels.

    90.44% accuracy, averaged over test folds 1 to 5 of the LUNA16 split (1,004 nodules, 450 positive). GBM on nodule size and raw pixels alone gave 86.12%, which the authors call comparable to a multi-scale CNN. The multi-scale CNN numbers are in Table 1 and are not given in the text. · Gradient Boosting Machine with 3D Dual Path Net…; DeepLung for Nodule Classification; Tables

  3. SuggestedOn each doctor's confident nodules, the classification subnetwork's average accuracy is above the average of four experienced doctors, and it beats three of the four.

    Average accuracy 1.5% above the doctors' average; better than doctors 1, 3 and 4 (Table 2). About 1/3 of each doctor's nodules were scored 3 and left out, and the authors note the remaining confident nodules are easier. · Compared with Experienced Doctors on Their…; Tables

  4. SuggestedIn the fully automated setting, with the classifier retrained on detected nodules, accuracy on correctly detected nodules stays above the doctors' average, and the classifier rejects nearly all false-positive detections.

    81.42% accuracy on detected true-positive nodules vs 78.36% average for doctors on all nodules; 97.02% of false-positive detected nodules removed by the classifier (Table 3, folds 1 to 5). · DeepLung for Fully Automated Lung CT Cancer…; Tables

  5. ShownAt patient level (a CT counts as positive if any detected nodule is called malignant), DeepLung performs about as well as the doctors on their confident CTs.

    81.41% accuracy, which is 99% of the four doctors' average and better than doctor 4 (Table 4); Kappa against ground-truth labels 63.02% vs 64.46% average for the doctors. · Compared with Experienced Doctors on Their…; DeepLung System: Fully Automated Lung CT Cancer…; Tables

Methods, data and reproducibility

Detection: a 3D Faster R-CNN (encoder derived from ResNet-18, U-net-like decoder, anchors of 5, 10 and 20) runs on 96×96×96 CT patches, followed by non-maximum suppression. Classification: 32×32×32 crops centred on detected nodules go through a 3D dual path network. Its features, plus nodule size and raw pixels, feed a gradient boosting machine. A CT is called positive if any of its nodules is.

  • LUNA16 low-dose lung CTs with detection annotations (nodules under 3mm removed) and the official patient-level 10-fold split · human (patient lung CT scans) · n = 888 CTs
  • LIDC-IDRI low-dose lung CTs with several doctors' annotations (size, location, diagnosis, texture, margin), restricted to LUNA16 CTs for classification · human (patient lung CT scans) · n = 1,018 CTs
  • Nodule malignancy set built from averaged LIDC-IDRI doctor scores (average of 3 removed; above 3 labelled positive) · human · n = 1,004 nodules (450 positive)
  • Diagnoses from the four doctors who labelled the most nodules, used as the human comparison · human · n = 4 doctors
Methods & tools
3D Faster R-CNN, U-net (encoder-decoder design), ResNet-18, 3D dual path network (DPN), Gradient boosting machine (GBM), Non-maximum suppression, FROC (LUNA16 evaluation metric), Kappa coefficient
Assay conditions & compute
Hardware is not specified. GPU memory limits set the 96×96×96 patch size and the batch size. Training used SGD for 150 epochs (detector) and 1,050 epochs (classifier).
Reproducibility gaps
No links to code, trained models or data; no hardware or software versions given. Classification and full-system results use only 5 of the 10 cross-validation folds.
Limitations (6)
  • Because of time and resource limits, classification and full-system results are averaged over test folds 1 to 5 only, not all 10 folds. · stated by the authors
  • The comparison with doctors uses each doctor's confident nodules, which the authors note are easier to diagnose; about 1/3 of each doctor's nodules were scored 3 and excluded. · stated by the authors
  • The remaining false-positive detections are very small and look nodule-like; the authors note that nodules under 3mm were removed from the ground truth. The classifier may miss weak irregular boundaries and call some malignant nodules benign. · stated by the authors
  • Detection and classification are trained separately rather than in one unified framework, and no non-imaging data (such as medical report history) is used; the authors list these as future work. · stated by the authors
  • Malignancy ground truth is the average of the doctors' own scores, with no independent confirmation described, so comparing the model with those same doctors is partly circular. · noted in this summary
  • All evaluation uses a single source (LUNA16/LIDC-IDRI) with no external cohort. The 81.42% vs 78.36% comparison also sets detected true-positive nodules against doctors' results on all nodules, which are different sets. · noted in this summary

Who should read:Read this if you build CT lesion detection or classification pipelines and want an early baseline that pairs a 3D region-proposal detector with gradient boosting on deep 3D features, benchmarked on LUNA16 and LIDC-IDRI against doctors.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISChen '20Zhu '18Bouziane '22Bizzego '19Wu '20Liu '26
76% semantic matchBioengineering
J. Chen, W. Wang, B. Ju+5 more·bioRxiv·Apr 14, 2020