Skip to main content

Table 8 Comparison of Different CLIP Adaptation Strategies on TCGA-based Cell Classification Dataset

From: Instance-level semantic segmentation of nuclei based on multimodal structure encoding

Method

\(F_{\text {c}\_\text {avg}}\)

\(F_{\text {immune}}\)

\(F_{\text {stroma}}\)

\(F_{\text {tumor}}\)

CLIP zero-shot

0.730

0.710

0.720

0.725

CLIP fine-tuning

0.785

0.765

0.775

0.780

CLIP linear probing

0.790

0.770

0.780

0.785

Ours (knowledge distillation)

0.858

0.822

0.875

0.877

  1. Our knowledge distillation approach shows superior performance compared to direct CLIP adaptation methods on TCGA-based Cell Classification Dataset