LungCraft: a hybrid 3D-2D deep learning and radiomics framework with explainable AI for precision diagnosis of lung cancer.
Authors
Affiliations (1)
Affiliations (1)
- School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India.
Abstract
Lung cancer is the top cause of cancer-related death, globally. The morphological complexity of tumours, intra-tumoural heterogeneity and limited interpretability of current imaging systems all contribute to the difficulty of early and reliable detection of tumours. This paper introduces LungCraft, a diagnostic framework composed of 3D medical modelling, quantitative radiomics and explainable artificial intelligence (XAI) for lung adenocarcinoma classification, using computed tomography (CT) scans as input. LungCraft uses a hybrid deep learning architecture (HybridNET) which has a combination of layers that use 3D convolutional layers for volumetric feature learning and 2D convolutional layers for contextual refinement. Sixty-one TCIA CT volumes were used for internal validation, while the LIDC-IDRI and NSCLC-Radiomics datasets were kept as held-out test sets for external validation. The spatial and feature level interpretability is given by Grad-CAM and KernelSHAP. Unlike previous methods which focussed on achieving one or more of these, LungCraft's main strength is that it brings these in a systematic way as part of an integrated pipeline that includes a two-stage feature-level fusion strategy, dual spatial and feature level explainability and interactive 3D tumour visualization. On internal test set, LungCraft was able to classify with an accuracy of 91.3% and an AUC of 0.93, whereas on external cohorts it was able to classify with an accuracy of 88.7% and an AUC of 0.91 (LIDC-IDRI) and 89.6% and 0.92 (NSCLC-Radiomics). 95% confidence intervals are given for all measurements. The results are statistically significant (<i>p</i> < 0.01) when compared with the results of the baseline architecture. LungCraft achieves accurate and interpretable diagnosis of lung adenocarcinoma, with demonstrated external generalisation and has a potential use as an AI assistant in clinical decision making which is subject to formal prospective validation.