Deep Learning for Lung Cancer Detection on Screening CT Scans: Results of a Large-Scale Public Competition and an Observer Study with 11 Radiologists
Bibliographic record
Abstract
Purpose To determine whether deep learning algorithms developed in a public competition could identify lung cancer on low-dose CT scans with a performance similar to that of radiologists. Materials and Methods In this retrospective study, a dataset consisting of 300 patient scans was used for model assessment; 150 patient scans were from the competition set and 150 were from an independent dataset. Both test datasets contained 50 cancer-positive scans and 100 cancer-negative scans. The reference standard was set by histopathologic examination for cancer-positive scans and imaging follow-up for at least 2 years for cancer-negative scans. The test datasets were applied to the three top-performing algorithms from the Kaggle Data Science Bowl 2017 public competition: grt123, Julian de Wit and Daniel Hammack (JWDH), and Aidence. Model outputs were compared with an observer study of 11 radiologists that assessed the same test datasets. Each scan was scored on a continuous scale by both the deep learning algorithms and the radiologists. Performance was measured using multireader, multicase receiver operating characteristic analysis. Results The area under the receiver operating characteristic curve (AUC) was 0.877 (95% CI: 0.842, 0.910) for grt123, 0.902 (95% CI: 0.871, 0.932) for JWDH, and 0.900 (95% CI: 0.870, 0.928) for Aidence. The average AUC of the radiologists was 0.917 (95% CI: 0.889, 0.945), which was significantly higher than grt123 (P = .02); however, no significant difference was found between the radiologists and JWDH (P = .29) or Aidence (P = .26). Conclusion Deep learning algorithms developed in a public competition for lung cancer detection in low-dose CT scans reached performance close to that of radiologists. Keywords: Lung, CT, Thorax, Screening, Oncology Supplemental material is available for this article. © RSNA, 2021
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.022 | 0.044 |
| Meta-epidemiology (narrow) | 0.002 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.002 | 0.001 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.001 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".