Finding lung nodules early can prevent the severe disability caused by advanced lung cancer. This AI looks at each nodule from three different views, much as a radiologist would, to decide whether it is harmless or cancerous.
The problem
Detecting lung nodules early can greatly improve survival, but it takes manual, laborious work by radiologists, with limited success. People with disabilities may also face unequal access to healthcare.
What the research did
We developed a Multi-View Soft Attention-Based Convolutional Neural Network (MVSA-CNN). It extracts patches of each nodule from three different views and uses “attention” to focus on the most informative parts. It classifies nodules as benign, primary cancer or metastatic (spread from elsewhere).
We trained and tested it on the public LIDC-IDRI lung CT database using 10-fold cross-validation.
What we found
MVSA-CNN outperformed competing methods with 97.10% accuracy, 96.31% sensitivity and 97.45% specificity.
Who benefits
- Radiologists, who gain a reliable assistant for nodule assessment
- Patients, through earlier, more reliable diagnosis
- People with disabilities, who may face barriers to specialist care
Limits and what comes next
Results are from one public dataset. The model would need testing on scans from different hospitals and populations, and clinical validation, before use.
The team
This research was led by Jannatul Ferdous Esha and colleagues at Bangladesh University of Professionals, with me, Mohammad Abu Yousuf, AKM Azad, Asmaa Soliman Al-Moisheer, Naif Alotaibi, Salem A. Alyami (Imam Mohammad Ibn Saud Islamic University) and Mohammad Ali Moni.
Frequently asked questions
What is a lung nodule?
A small spot in the lung seen on a CT scan. Most are harmless, but some are cancer.
Why look from three views?
Different views reveal different features of the nodule, making classification more reliable.
How accurate is MVSA-CNN?
97.10% accuracy on the LIDC-IDRI dataset.