Pathology image, radiomics, spatial transcriptomics(ST)๋ฅผ ํตํฉํ๊ธฐ ์ํ ์ฐ๊ตฌ ์ํฌ์คํ์ด์ค์ ๋๋ค.
Pathology image๋ก๋ถํฐ ST๋ฅผ ์์ธกํ๋ ๊ณผ์ ์ radiomics๋ฅผ ๊ฒฐํฉํ๋ ๋ ๊ฐ์ง ํ๋ ์์ํฌ๋ฅผ ์๊ฐํฉ๋๋ค.
Radiomics-Pathomics Contrastive Learning for Interpretable ST Prediction
RaPaCL์ ๋ณ๋ฆฌ ์ด๋ฏธ์ง ๊ธฐ๋ฐ ST ์์ธก ๋ชจ๋ธ์ ํด์ ๊ฐ๋ฅํ radiomics ํํ์ ํตํฉํ๊ธฐ ์ํ ํ๋ ์์ํฌ์ ๋๋ค. Cell-segmented patch์์ ์ถ์ถํ handcrafted radiomics feature์ ๋ฅ๋ฌ๋ ๊ธฐ๋ฐ pathomics feature๋ฅผ contrastive learning์ผ๋ก ๋์ผํ latent space์ ์ ๋ ฌํ์ฌ, ๋ชจ๋ธ์ด ์กฐ์ง์ textureยทshapeยทmorphology์ ํต๊ณ์ ์์น์ ๊ฐ์ ํด์ ๊ฐ๋ฅํ ํน์ฑ์ ํ์ตํ๋๋ก ์ ๋ํฉ๋๋ค. ์ ๋ ฌ๋ radiomics-pathomics ํํ์ ์ตํฉํ์ฌ spot-wise gene expression์ ์์ธกํ๋ฉฐ, ์์ธก ์ฑ๋ฅ๊ณผ ํด์ ๊ฐ๋ฅ์ฑ์ ๋์์ ํฅ์์ํค๋ ๊ฒ์ ๋ชฉํ๋ก ํฉ๋๋ค.
Spatially-aware Radiomics Network for ST Prediction
STaRN์ ๊ฒฝ๋ radiomics ํํ์ ํ์ฉํ์ฌ ๊ณต๊ฐ์ ์ด์ ์ ๋ณด(spatial context) ๋ฅผ ํจ๊ณผ์ ์ผ๋ก ํ์ตํ๋ ํ๋ ์์ํฌ์ ๋๋ค. ๊ฐ spot๊ณผ ์ฃผ๋ณ ์ด์ spot๋ค์ Summary Table ํํ๋ก ๊ตฌ์ฑํ ๋ค, SAINT ๊ธฐ๋ฐ dual-attention encoder๋ฅผ ์ฌ์ฉํ์ฌ spot ๋ด feature ์ํธ์์ฉ(column attention)๊ณผ ์ด์ ๊ฐ ๊ด๊ณ(row attention)๋ฅผ ๋์์ ๋ชจ๋ธ๋งํฉ๋๋ค. ๋ํ UNI ๋ฐ scFoundation ๊ธฐ๋ฐ teacher representation์ผ๋ก๋ถํฐ distillation์ ์ํํ๊ณ , self-contrastive learning์ ๊ฒฐํฉํ์ฌ ๊ณต๊ฐ์ ยท์๋ฏธ์ ๋งฅ๋ฝ์ ๋ฐ์ํ๋ radiomics ํํ์ ํ์ตํฉ๋๋ค. ํ์ต๋ ํํ์ downstream gene expression prediction์ ํ์ฉ๋ฉ๋๋ค.

