Healthcare & Bio AI
New pathology slide classifier outperformed recent methods on two datasets
In a 2022 paper presented at the IEEE/CVF Conference on Computer Vision and Pattern Recognition, researchers from the University of Liverpool and the Chinese Academy of Sciences report that a method called DTFD-MIL classified whole slide images more accurately than other recent methods on the CAMELYON-16 benchmark and on a lung cancer dataset from TCGA.
Whole slide images are large digital scans of tissue samples used in pathology. Multiple instance learning is a way to train a classifier when labels apply to groups rather than to each small patch. Here each slide is a group, or bag, and the many patches cropped from it are the individual examples. The authors write that small sample cohorts create a challenge. There are few slides available, but each high-resolution slide produces a large number of patches.
How does DTFD-MIL work?
DTFD-MIL introduces pseudo-bags to virtually enlarge the number of slide groups. It then uses a two-level, double-tier multiple instance learning design to make better use of the features already present. The researchers also derive an instance probability within attention-based multiple instance learning, a technique that assigns different weight to patches, and use that derivation to help build and analyze the method.
What did the evaluation show?
The authors tested the method on the CAMELYON-16 dataset and the TCGA lung cancer dataset. On CAMELYON-16, they report that DTFD-MIL outperformed the latest comparison methods by a large margin. It also did better on the TCGA lung cancer dataset. The authors say the framework is ready to be extended to wider multiple instance learning applications.
What else should a reader know?
What is a whole slide image?
A whole slide image is a large digital scan of a tissue sample used in pathology.
What are pseudo-bags?
Pseudo-bags are virtual groups the method creates to enlarge the number of slide groups when few slides are available.
Which datasets did the study use?
The study evaluated on CAMELYON-16 and the TCGA lung cancer dataset.
Limits and what's next
The results come from two benchmark datasets, the CAMELYON-16 dataset and the TCGA lung cancer dataset. The evaluation is a benchmark comparison, not a report from clinical use. This is one study, and the abstract reports no independent replication. The authors describe the framework as ready to be extended to wider applications, but the abstract does not report results from those extensions.
Source: Hongrun Zhang et al., “DTFD-MIL: Double-Tier Feature Distillation Multiple Instance Learning for Histopathology Whole Slide Image Classification,” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.