Accurate subtyping of lung cancers by modelling class dependencies
Identifying subtypes and histological patterns is crucial for lung cancer diagnosis and treatment. Nevertheless, datasets with complete subtyping annotations are scarce, and most existing work primarily focuses on categorising lung cancers into fundamental types, omitting the distinction of adenocar...
| Main Authors: | , , , , , , |
|---|---|
| Format: | Conference item |
| Language: | English |
| Published: |
IEEE
2024
|
| Summary: | Identifying subtypes and histological patterns is crucial for
lung cancer diagnosis and treatment. Nevertheless, datasets
with complete subtyping annotations are scarce, and most
existing work primarily focuses on categorising lung cancers into fundamental types, omitting the distinction of
adenocarcinoma patterns. We present a computational approach for a more comprehensive lung cancer subtyping
from histology by modelling the dependencies between
cancer subtypes and histological patterns in a multi-label
setting. Our approach utilises slide-level labels indicating cancer subtypes as well as the presence of cancerassociated patterns, thereby alleviating the need for labourintensive region-based annotations. A new dataset with
cancer-associated pattern labels is constructed and combined
with publicly available datasets. We evaluate our model’s
ability to simultaneously differentiate cancer subtypes and
cancer-associated patterns. The result demonstrates that
our modules enable conventional weakly-supervised classification models on multi-label problems, achieving subset
accuracy of 84% when differentiating lung cancer subtypes
and cancer-associated histological patterns. |
|---|