Multiple Types of Cancer Classification Using CT/MRI Images Based on Learning Without Forgetting Powered Deep Learning Models
Cancer is one of the world’s leading causes of death. AI-assisted screening can help evaluate more cases in less time — if the model can hold onto everything it has learned.
Venue
IEEE Access, vol. 11
Pages / Year
10336–10354 · 2023
Role
Co-author · undergrad research
DOI
10.1109/ACCESS.2023.3240443
01 · Research problem
IndexChapter 01
The research problem
Early detection significantly improves survival, and AI-assisted screening could evaluate far more cases than human review alone. But a single model that handles eight different cancers faces the same trap as the fire work: transfer learning onto a new cancer dataset can destroy the model’s ability to classify the datasets it was originally trained on.
Chapter 02
Method
In plain terms: one model screens images for eight different cancers, tuned automatically rather than by hand — without losing old skills as it learns new ones.
Stage 01 — CT/MRI imagery
Images carrying cancer traits across eight kinds of cancer — including lung, brain, breast, and cervical — drawn from the companion Multi Cancer Dataset.
Illustrative pipeline — the paper reports the exact architectures and search space.
Chapter 03
Results, as published
The paper reports that the proposed transfer-learning models are more accurate than the then state-of-the-art techniques, and that LwF classifies both new datasets and previously trained datasets better than the alternatives. The article spans 19 pages in IEEE Access volume 11 and contributes to UN Sustainable Development Goal 3 (Good Health and Well-being).
Chapter 04
Scope and authorship
Cite this
Subramanian, M., Cho, J., Sathishkumar, V. E., & Naren, O. S. (2023). Multiple Types of Cancer Classification Using CT/MRI Images Based on Learning Without Forgetting Powered Deep Learning Models. IEEE Access, 11, 10336–10354.
doi.org/10.1109/ACCESS.2023.3240443 ↗Lessons
What it taught me
- Publishing the dataset alongside the paper multiplied its reach far beyond what the paper alone achieved — the Kaggle corpus has since seeded well over a hundred community notebooks.
- Bayesian optimisation over hand-tuning is the same instinct as A/B testing over opinion: let the search tell you, not the loudest engineer.
- Medical machine learning has a duty of humility. A classifier is a screening aid, never a diagnosis.