Multimodal AI for Pulmonary Nodule Classification: A Review

The classification of pulmonary nodules by computed tomography (CT) is a crucial step in computer-aided lung cancer diagnosis, but it is still time-consuming and highly inter-observer dependent, and pertinent literature of convolutional neural networks (CNN) for this problem is extensive, diverse, and hard to synthesize. This paper presents a systematic review and categorizes the CNN related pulmonary nodule classification studies into an eight-stage pipeline with an additional stage of explainable AI, in order to make it easier to refer to different studies: Dataset curation, CT preprocessing, ROI extraction, CNN backbone selection and ensembling, attention based feature fusion, radiomic and clinical feature integration, multimodal classification, uncertainty quantification, explainable AI, and clinical validation. The backbone architectures and attention-based fusion are fairly well developed, whereas uncertainty quantification, validation of explainability, and eventual clinical utility are less mature, with pooled external-validation studies reporting greater than 85% sensitivity and less than 80% specificity. There are nine identified research gaps that range from diversity of datasets to harmonization of preprocessing, benchmarking in fusion, and outcome-related validation. The results provide a useful guide for researchers and clinical-AI developers looking for pulmonary nodule classification systems that can extrapolate from the benchmark set and have clinical confidence.