A team of clinicians and engineers is building an artificial intelligence system to spot lung cancer earlier, by training it on scans from over 1,500 patients with suspicious lung nodules. Lung cancer kills more people than any other cancer, partly because it is often caught too late. Current methods for assessing small lung nodules—which could be early cancers—rely on size and growth rate, but miss many aggressive tumours. This project aims to close that gap by teaching a machine-learning algorithm to recognise subtle texture patterns in CT scans that indicate malignancy, even in nodules too small to measure reliably. If the system works, it could be deployed directly into NHS hospitals as a CE-marked medical device. Radiologists would receive a risk score for each nodule, helping them decide which patients need urgent follow-up and which can be safely monitored. That could reduce unnecessary biopsies, shorten diagnostic delays, and improve survival rates. The researchers are also building the database to be reusable for other lung conditions, such as COPD, so the infrastructure could support future diagnostic tools beyond cancer.
View original technical description
IDEAL comprises 6 work packages (WP) over 36 months. Each WP has a leader, with Milestones, Tasks (T) and Deliverables (D). See Appendix A for a GANTT chart. WP1: Project Management (M1-M36)(Lead: FG, Involved: VP-DB-MC-SMB, FTE:1.35) The project will be led by FG supported by a Project Manager (PM), who will ensure that the project meets the required quality, time & cost, identify risks, manage spend & provide reports. D1.1. Risk assessment (M1)D1.2. Regular progress reports WP2: Planning for a Big Dataset (M1-M9)(Lead: FG, Involved: LP-MC-DB, FTE:3.75) T2.1. Database Definition (M1)We will define the database architecture to enable selection of data subsets to adaptively re-train a best classifier for the patient demographic/scanner/pathway. Project partners will test the baseline system and decide what annotations will be included, including making the database multi-use (e.g. including COPD annotations). We will define thresholds for minimum acceptable performance. T2.2. Big Data Acquisition and Testing (M2-M9)We will collect >500 patients containing pulmonary nodules (PN) with ground-truth per site (total: 1500). The PN Clinics at three sites see 90 patients per month and have a database of over 1000 patients. The new database will be used to augment pre-existing datasets. The database creation will require an innovative data science approach, so that the thousands of images represent the millions of possible patient, scanner, and protocol combinations. The clinical research fellow and radiographers will collect, annotate and anonymize data. They will iteratively re-train and test the algorithms in collaboration with Optellum, in order to identify nodule types (e.g. sizes, characteristics, clinical pathways) and scan types (scanner model, protocol, kernel, resolution) to be added in order to create a balanced and unbiased set. Milestone 1: Database defined (M1)Milestone 2: Database ready (M9)D2.1. New PN database with 1500 patients WP3: Product Development (M1-M15)(Lead: VP, Involved: LP-SMB, FTE:4.625) T3.1. Machine Learning on Multi-Center Data (M1-M9, FTE:1.5)The evolving database from WP2 will be integrated into a “lean product development” to improve machine learning and guide data acquisition, so that the DIB handles heterogeneous data. Optellum will productize the LVB prototype so that it can be robustly deployed in a hospital setting. T3.2. Front-end Prototype (M1-M3)(FTE:0.5)User interface prototype defined and developed, with frequent Usability Tests with the end users. T3.3. Back-end Prototype(M3-M9)(FTE:0.5)Back-end server ready for PACS integration developed and tested. T3.4. Final Product Development (M10-M15)(FTE:1.5) T3.5. CE-marking (M1-M15)(FTE:0.625)We determine regulatory classification according to CE-MDD. Quality Management System (QMS) will be introduced and user testing conducted with help of consultant from ScreenPoint Medical. We expect to apply for CE-marking in M12. Milestone 3: Prototype ready for clinical study (M9) Milestone 4: CE-marking (M15) D3.1. Full software D3.2. CE-mark WP4: Texture Risk-model (M1-M12)(Lead: FG, Involved: LP-MC-DB-SMB, FTE:1.0) T4.1. Risk-model development. We will develop a risk model for small pulmonary nodules that incorporates texture using retrospective data. Milestone 5: Texture included in risk model (M9) D4.1. Risk model report (M12) WP5: Clinical Study (M9-M33)(Lead FG, Involved: MC-DB, FTE:10.0) T5.1. Prospective clinical study (M9-M33) Research radiographers at the three sites, under the guidance of the clinical research fellow, will assess patients with PN and enter them into the clinical trial and follow them for up to 12 months with established methods (volumetry, PET) compared against the initial DIB score. Milestone 6: Final patient follow up (M33) D5.1. Study results report (M36) WP6: Analysis & Dissemination (M21-M36)(Lead FG, Involved: MC-DB-VP, FTE:1.25) T6.1. NHS dissemination: Results presented at BTS and BSTI meetings. T
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know