label-free bright-field images and models about NSCLC

This repository provides label-free bright-field image datasets of patient-derived non-small cell lung cancer (NSCLC) organoids, along with deep learning models for subtype classification and drug sensitivity evaluation. The datasets include (i) untreated organoids for distinguishing lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), and (ii) cisplatin-treated organoids with multi-time-point imaging for early-stage drug response analysis.

We further provide two models: IMANet, an interpretable morphological attention network integrating multi-scale attention with SHAP-guided feature selection for subtype classification; and MT-MANet, a multi-task framework with a shared attention backbone and dual task-specific heads for joint subtype recognition and drug response prediction.

The datasets comprise 1,956 images from 9 patients (15 batches) for subtype classification, and 18,641 images across multiple time points for drug response evaluation. Experimental results demonstrate strong subtype discrimination (AUC = 0.940) and improved performance in the multi-task setting (AUC = 0.995), as well as effective early-stage drug response prediction.