Robust artificial intelligence frameworks for lung cancer subtyping and malignancy detection on thoracic CT

Description

Mustafa Tahsin Yilmaz, Enes Algul, and Ishak Pacal benchmark, in a 2026 comparative study, thirty deep-learning architectures for two lung cancer tasks on chest CT scans. Sorting tumors by tissue subtype and judging nodule malignancy need different visual cues, one broad and contextual, the other local. Fifteen convolutional networks and fifteen vision transformers were trained under one unified setup.

Two public datasets supplied images, one covering three cancer subtypes plus normal scans and one covering normal, benign, and malignant cases. A transformer model led subtyping with an F1 score near 0.973, closely followed by a leading convolutional model. For malignancy classification, three models tied at the top near 0.982. Heat-map visualizations suggested the best models attended to lesion and lung regions that matter clinically.

Catalog record

Source: https://europepmc.org/article/MED/PMC13527102