The diagnostic gap in lung cancer: Most lung cancers are diagnosed at advanced stages when survival rates drop dramatically. Two promising early detection approaches - miRNA biomarkers in blood and CT radiomics - each have limitations individually. Combining them through deep learning multimodal integration could bridge this gap.
What this review covers: This comprehensive review paper covers the biology of miRNA biomarkers in lung cancer, nanomaterial-based biosensor technologies for detecting them, radiomics feature extraction from CT imaging, and deep learning architectures that integrate both data types for diagnosis.
Why multimodal integration matters: miRNA signatures capture molecular-level tumor biology (gene expression dysregulation), while radiomics captures tumor morphology and texture from imaging. These two information sources are complementary - each captures what the other misses.
The fusion approach: The DenseNet-based multimodal fusion architecture highlighted in this review achieved AUC of 0.98 for lung cancer diagnosis - significantly higher than either miRNA biomarkers alone (AUC approximately 0.85) or radiomics alone (AUC approximately 0.88).
What are miRNAs: MicroRNAs (miRNAs) are small non-coding RNA molecules that regulate gene expression by binding to messenger RNAs and suppressing their translation. They are produced in cells and can be detected in blood plasma or serum, where they circulate in stable forms protected inside exosomes or bound to proteins.
Lung cancer-specific miRNA signatures: Multiple miRNAs are consistently dysregulated in lung cancer. miR-21, miR-210, and miR-155 are commonly upregulated, while tumor-suppressive miRNAs like miR-34a and let-7 family members are downregulated. Panels of multiple miRNAs improve diagnostic specificity over single markers.
miRNAs vs. ctDNA: Compared to circulating tumor DNA (ctDNA), miRNAs offer certain advantages: they are more abundant in blood, more stable, and can reflect not just tumor mutations but also the tumor's interaction with the immune microenvironment. However, ctDNA provides more specific genetic information about driver mutations.
Technical challenges: miRNA detection faces challenges of low concentration, susceptibility to hemolysis (red blood cell lysis), and lack of standardized extraction protocols. These issues affect reproducibility across different laboratories and detection platforms.
Why nanomaterial biosensors: Traditional miRNA detection methods like qPCR require expensive equipment and complex sample processing. Nanomaterial-based biosensors offer an alternative: highly sensitive, potentially point-of-care detection platforms that can detect miRNAs at very low concentrations with minimal sample preparation.
Gold nanocluster biosensors: Gold nanoclusters functionalized with DNA probes complementary to target miRNAs can detect specific miRNA sequences through fluorescence or electrochemical signals. Their large surface area and tunable optical properties make them extremely sensitive detection platforms.
Metal-organic framework (MOF) sensors: MOFs are porous crystalline materials with enormous surface areas. When miRNAs bind to MOF-integrated probes, they change the material's electrical or optical properties in measurable ways. MOF sensors achieve femtomolar detection limits - far beyond what traditional methods can achieve.
Quantum dot-based detection: Quantum dots are semiconductor nanocrystals that emit light at specific wavelengths when excited. Conjugated with miRNA-specific probes, they enable multiplexed detection of multiple miRNAs simultaneously from a single blood sample, providing richer diagnostic information.
What is radiomics: Radiomics is the process of extracting large numbers of quantitative features from medical images - features invisible to the human eye - that capture tumor texture, shape, intensity patterns, and heterogeneity. From a single CT scan, thousands of radiomics features can be extracted.
Standard radiomics feature categories: Features include first-order statistics (intensity distributions), shape descriptors (tumor volume, compactness), and texture matrices (GLCM for local intensity variations, GLRLM for run-length patterns). Wavelet transformations extract features at multiple frequency scales.
Feature selection challenge: With thousands of features from a small number of patients, overfitting is a major risk. Dimensionality reduction methods like LASSO regression, principal component analysis (PCA), and tree-based feature importance are essential to identify the handful of features that genuinely predict lung cancer presence or subtype.
Deep learning radiomics: Rather than manually crafted features, convolutional neural networks (CNNs) applied directly to CT images can automatically learn relevant spatial patterns. These deep features often capture complementary information to hand-crafted radiomics, and combining both approaches improves performance.
The fusion challenge: miRNA expression data and CT radiomics are fundamentally different data types - one is a high-dimensional molecular profile, the other a set of image-derived quantitative features. Fusing them requires a model architecture that can process each modality appropriately before combining their representations.
DenseNet-based fusion: The DenseNet architecture, which uses dense connections between layers, is particularly effective because each layer receives input from all previous layers. Applied to both miRNA profiles and radiomics features in parallel branches, DenseNet extracts highly informative representations from each modality.
Late vs. early fusion: The review describes both early fusion (concatenating raw features before the model) and late fusion (training separate models for each modality and combining their predictions). Late fusion tends to perform better when the modalities have very different characteristics and noise profiles.
Diagnostic performance: The combined multimodal DenseNet fusion model achieved AUC of 0.98 for lung cancer diagnosis, representing a substantial improvement over individual modalities. Sensitivity and specificity both exceeded 93%, approaching what is needed for practical clinical deployment.
Data standardization: A major barrier to clinical translation is lack of standardized protocols for both miRNA isolation from blood and CT image acquisition. Model performance degrades when applied to data from different laboratories or scanner manufacturers.
Prospective clinical validation: All current results come from retrospective analyses of existing datasets. Prospective studies in screening populations - where lung cancer prevalence is low - are essential to characterize the model's true false-positive rate, which determines its clinical utility.
Regulatory pathway: AI diagnostic devices require regulatory approval (FDA 510(k), CE marking). The dynamic nature of deep learning models - which may be retrained as new data becomes available - creates regulatory challenges for ensuring consistent, auditable performance.
Cost-effectiveness analysis: For widespread deployment, multimodal testing needs to demonstrate cost-effectiveness compared to low-dose CT screening alone. This requires health economics modeling incorporating test costs, downstream management, and survival outcomes.
Expanding to additional modalities: Future multimodal systems should incorporate additional data types: protein biomarkers (CYFRA 21-1, CEA), clinical risk factors (smoking history, spirometry), and PET metabolic imaging. Each additional modality can potentially improve diagnostic confidence.
Subtype classification: Beyond distinguishing cancer from benign lesions, multimodal models could classify cancer subtype (adenocarcinoma vs. squamous cell vs. small cell), predict molecular status (EGFR, KRAS mutations), and estimate prognosis - enabling simultaneous detection and molecular profiling.
Federated learning for multi-site training: Training robust models requires diverse data from multiple institutions. Federated learning - where models are trained locally at each site without sharing patient data - enables multi-site collaboration while preserving privacy.
Point-of-care integration: The ultimate vision is integration of nanomaterial biosensor-based blood testing with portable CT or chest X-ray at the point of care - enabling community-level lung cancer screening in resource-limited settings, particularly in high-prevalence populations.