Exploring Fairness and Performance Drivers Across State-of-the-Art Pulmonary Nodule Detection Algorithms
Bibliographic record
Abstract
Lung cancer is the leading cause of cancer-related deaths in the UK. Its high mortality rate is primarily due to its asymptomatic nature in the early stages, leading to late-stage diagnoses. However, effective early detection methods, such as Low-Dose Computed Tomography (LDCT), and treatments for early-stage disease make lung cancer an ideal candidate for screening. The UK Government aims to implement a national lung cancer screening programme targeting high-risk populations by 2029. This will significantly increase the workload on an already stretched radiology workforce, driving the adoption of computer-aided detection (CADe) systems to support radiologists. The datasets used to train these algorithms are typically drawn from previous lung cancer screening trials and studies (National Lung Screening Trial Research Team (2011); de Koning (2020)), which often lack balanced representation of protected groups, such as sex and ethnicity. This project examines whether training nodule detection algorithms on low-dose computed tomography (LDCT) scans from a London-based lung screening study, where these groups are typically under-represented, affects algorithm performance for under-represented categories. Our results indicate that overall performance remains equitable across all categories, even when trained on unbalanced datasets. The discriminative performance of deep learning-based pulmonary nodule detection algorithms is primarily driven by the composition of the dataset, specifically, the relative proportion of nodule types and sizes, rather than by protected attributes such as sex or ethnic group. The features learned from the nodules themselves drive detection outcomes, meaning that in populations where the prevalent nodule characteristics closely match the training data, performance is likely to be strong. While this study found no demographic disparities for nodule detection, there is no guarantee that this will be true across all populations, particularly those in populations where cancer risk predominates within different nodule distributions. This study provides an early assessment of performance variations of deep learning models across under-represented groups within a standard lung cancer screening dataset. While previous research has focused on improving how well nodule detection algorithms identify pulmonary nodules, this study uniquely focuses on demographic performance disparities and the impact of training data composition and algorithm design on model generalisability. The findings highlight critical considerations for the deployment of CADe systems in lung cancer screening, ensuring equitable performance across diverse patient populations. Our code is available at https://github.com/johnmccabe44/fairness-in-nodule-detection
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 | 0.030 | 0.108 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.004 | 0.005 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.002 | 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".