WCLC 2025 – AI subclassification of abnormal cells into pre-cancerous and malignant categories based on 3D cell imaging

Control # 2025-RA-1371-WCLC

Meyer, Wilbur, Nelson

Introduction

VisionGate’s Cell-CT™ platform computes 3D images of cells with isometric sub-micron resolution and measures >500 orientation-invariant 3D features in each individual cell. The LuCED® assay uses home-collected sputum to detect early-stage lung cancer through AI discovery of abnormal cells across the cytologic categories of atypia through to malignancy. Results show that early-stage lung cancer is detected with sensitivity and specificity both exceeding 90%, and the sensitivity to biopsy-confirmed stage 1 NSCLC lung cancer was 93.8%. In this context, we hypothesize that the abnormal cell population may be sub-classified into malignant and pre-cancer categories. This application will broaden the clinical scope of the LuCED assay to prioritize patient triage in cases where frankly cancerous cells are detected. Cases without malignant cells would default to standard protocols of vigilance with a likelihood of pre-cancer.

Methods

Abnormal cells detected with LuCED were further classified into cancer and pre-cancer categories using an AI-supervised learning process involving cytologically defined cells in classes of pre-cancer (N=623) and malignancy (N=374). The Cell-CT measured 503 3D features for each cell and adaptively boosted logistic regression separated the two classes. Ten-fold cross-validation ensured that the resulting classifier was not overtrained.

Results

A model involving six features emerged from the training process, producing an area under the ROC of 0.81. A potential operating point produced these results: 70%

detection of malignant cells, with 30% of malignant cells called pre-cancer; 80% detection of pre-cancer cells, with 20% of pre-cancer cells called malignant; None of the abnormal cells were called normal.

Conclusions

These results demonstrate the efficacy of 3D cell analysis in characterizing subtle aspects of abnormal cell morphology to assess whether a malignant process is likely to be present. This new AI classifier may find application in the COPD patient population and in improving the management of IPNs. In such populations, a clear indication of malignancy is urgently needed to initiate curative/lifesaving therapy, particularly in early-stage disease, where such treatments will be most effective.