Why Nodule Type Matters. Pulmonary ground-glass opacity (GGO) nodules detected on CT scans represent a diverse spectrum of pathology, from benign inflammation to early-stage lung cancer. GGO can be classified into pure ground-glass opacity (PGGO), mixed ground-glass opacity (MGGO), solid nodules (SN), special types, and normal tissue -- each carrying distinct malignancy risk, management guidelines, and surgical timing recommendations.
Limitations of Visual Diagnosis. In clinical practice, most lung nodule classification is performed by radiologists through CT scan visual inspection. This approach is subjective, time-consuming, and highly dependent on physician experience, leading to variability in nodule management across institutions and individual practitioners. As LDCT screening becomes more widespread, the volume of nodules requiring evaluation is increasing rapidly.
Deep Learning Opportunity. Convolutional neural networks (CNNs) have demonstrated strong performance in image recognition across many domains, offering automatic feature learning, translational invariance, and scalability to large datasets without manually engineered features. Applying CNN-based methods to lung nodule classification promises to reduce diagnostic variability and improve classification accuracy beyond traditional visual assessment.
The Multi-Scale Problem. A fundamental challenge in nodule classification is scale variability: the same nodule type may appear at very different sizes across patients, and single-scale CNN models struggle to capture both fine textural detail and broader spatial context simultaneously. This study proposes integrating Gaussian Pyramid Decomposition (GPD) with CNN architecture to address this scale challenge through a novel multiscale approach.
Gaussian Pyramid Decomposition. Gaussian Pyramid Decomposition (GPD) generates multiple resolution versions of each input image through iterative Gaussian smoothing and downsampling. The Difference of Gaussians (DOG) variant extracts scale-space information by computing differences between adjacent pyramid levels, enabling detection of features across a wide range of spatial scales. This approach provides scale invariance -- the ability to detect nodule features regardless of size variation -- which standard CNN convolutions cannot achieve.
MCNN Architecture Design. The proposed Multiscale Convolutional Neural Network (MCNN) takes as input four multiscale images derived from GPD: the original 640 x 640 pixel image and three Gaussian difference pyramid levels. A slice layer separates these into independent channels for initial processing through the first convolutional layer (96 filters, 9 x 9 kernels). After pooling, features from all four scales are concatenated into a unified 384-channel multiscale representation, which is processed by two additional convolutional layers and three fully connected layers to produce five-class classification.
Architecture Details. The nine-layer network uses ReLU activation throughout, which avoids the gradient vanishing problems of sigmoid and tanh functions while enabling fast training. Dropout at rate 0.5 in the two largest fully connected layers (each with 4096 neurons) prevents overfitting on the medical imaging dataset. The final output layer has five neurons corresponding to the five nodule classification categories: solid nodules (SN), pure GGO (PGGO), mixed GGO (MGGO), special type, and normal.
Classifier Comparison Strategy. Two classification approaches were compared: MCNN with end-to-end Softmax classification, and MCNN with support vector machine (SVM) classification using features extracted from the pretrained MCNN. For the SVM approach, five binary one-versus-rest SVM classifiers were trained separately, with final prediction determined by the highest confidence score. This comparison isolates the effect of classifier choice from the multiscale feature extraction contribution.
Five Nodule Types. The classification task targeted five categories derived from established clinical criteria: solid nodules (SN, CT value 50-70 HU, diameter less than 0.3 cm mediastinal septum), pure ground-glass opacity (PGGO, CT value -600 to -700 HU, 0 to 10 nodules allowed), mixed ground-glass opacity (MGGO, CT value -400 to -500 HU, diameter greater than 0.3 cm), special types, and normal. Each category has distinct CT density, size, and structural characteristics that define the classification criteria.
Dataset Construction. Over 4,000 raw lung nodule images were collected from two sources: clinical patients at Chongming Hospital (Shanghai University of Medicine and Health Sciences) and The Cancer Imaging Archive (TCIA) public database. Images were normalized and standardized to 640 x 640 pixels, then augmented by rotation to produce over 10,000 standard sample images. The dataset was split into training (1,800 per class, total 9,000), validation (200 per class, total 1,000), and test (200 per class, total 1,000) sets with balanced representation across all five classes.
Preprocessing with GPD. Each standardized 640 x 640 image underwent multi-scale preprocessing through GPD to generate four different scale versions. This created a multiresolution training dataset in which each patient's nodule was represented across all relevant spatial scales simultaneously. The GPD components were implemented from the SIFT framework, utilizing only the pyramid construction and DOG computation stages without keypoint detection or descriptor generation.
Training Configuration. Both MCNN variants shared identical training parameters: batch size of 32, learning rate of 0.001 with stepwise reduction policy, weight decay of 0.0005, momentum of 0.9, and maximum 2,000 iterations with early stopping based on validation performance. The Caffe deep learning framework was used for all experiments. Training required approximately 18 hours on an NVIDIA GeForce RTX 3080 GPU with 10GB memory.
MCNN + Softmax Achieves Highest Accuracy. The MCNN combined with Softmax classifier achieved overall classification accuracy of 96.48%, outperforming both MCNN + SVM (94.21%) and standard CNN + Softmax (92.34%). For the normal class, 100% precision was achieved by both MCNN variants, confirming strong ability to exclude non-nodule tissue without false positives.
Softmax Outperforms SVM as Classifier. Comparing the two MCNN classifier variants, Softmax consistently outperformed SVM across all five nodule types, with F1 score improvements of 3.23 to 5.28%. PGGO nodules showed the largest improvement (5.28% higher F1 with Softmax), followed by MGGO (5.18%). This advantage arises because Softmax simultaneously optimizes all classes in a unified probabilistic framework, while SVM's one-versus-rest decomposition creates inconsistent class boundaries from separate binary classifiers.
Multiscale Features Improve Over Single-Scale CNN. Comparing MCNN + Softmax versus standard CNN + Softmax isolates the contribution of Gaussian Pyramid Decomposition. MCNN improved F1 scores by more than 2.27% for solid nodules, PGGO, and normal types. Overall accuracy improved from 92.34% (standard CNN) to 96.48% (MCNN), a 4.14 percentage point gain attributable entirely to multiscale feature extraction. The improvement was largest for nodule types with the greatest size variability.
Competitive with Prior Literature. Prior methods for lung nodule detection including 3D Faster R-CNN plus CMixNet achieved 96.33% accuracy but only for binary benign-malignant classification. Other multi-class approaches reached maximum accuracy of 95.5% and F1 of 95.24%. The MCNN achieved 96.48% accuracy and F1 scores exceeding these benchmarks while performing five-class classification, a substantially more challenging task than binary detection.
Mathematical Foundation of Scale Invariance. Standard CNN convolutions use fixed kernel sizes optimized for one spatial scale, limiting their ability to recognize the same nodule type at different sizes. Gaussian Pyramid Decomposition provides a mathematically grounded scale-space representation that simultaneously captures fine textural details at the original resolution and broader contextual patterns at coarser scales. This deterministic preprocessing approach achieves competitive results with significantly reduced computational complexity compared to attention-based multiscale methods like Swin Transformers.
Advantages Over Architectural Multiscale Approaches. Feature Pyramid Networks, U-Net architectures, and attention-based approaches achieve multiscale processing through complex network designs requiring attention parameter tuning and substantial computational resources during both training and inference. The GPD-CNN combination provides four distinct advantages: mathematical scale-space foundation, computational efficiency without attention computations, deterministic consistent processing, and architectural simplicity with standard CNN structures enhanced by preprocessing. This makes it especially suitable for resource-constrained clinical environments.
Proposed Clinical Workflow Integration. The MCNN can be integrated into existing clinical workflows as a computer-aided detection and diagnosis (CAD) tool operating alongside radiologist review. The proposed workflow includes: automatic GPD preprocessing of incoming CT scans, real-time classification with confidence scores, generation of structured reports by risk category, and presentation through existing PACS interfaces. Inference time averages 0.3 seconds per image, supporting real-time integration with routine radiology workflows.
Clinical Value for Nodule Management. The ability to distinguish PGGO, MGGO, and SN categories provides actionable surgical timing guidance: PGGO less than 5 mm are typically benign requiring only surveillance, while MGGO and SN with higher malignancy potential may require more immediate workup. The Fleischner Society recommends PET-CT for partially solid GGO lesions of 8-10 mm, and the MCNN's classification of MGGO versus PGGO can directly guide which patients require this additional imaging.
Model Contributions. The MCNN integrating Gaussian Pyramid Decomposition achieves 96.48% five-class lung nodule classification accuracy, outperforming traditional single-scale CNN models by over 4 percentage points. The multiscale approach specifically benefits nodule types with the greatest size variability, and the Softmax classifier's unified probabilistic framework outperforms one-versus-rest SVM decomposition by up to 5.28% F1 per class.
Key Limitations. The dataset does not include all clinically encountered nodule pathologies -- hamartomas, bronchial adenomas, and papillomas were excluded -- limiting applicability to the full spectrum of lung nodule types. The dataset combines images from a single hospital and public TCIA data with potentially different scanner protocols, introducing inter-scanner variability. The 18-hour training requirement and 6GB inference memory may exceed standard clinical workstation capacity in resource-limited settings.
Absence of Longitudinal Data. A significant limitation is that the model was trained and evaluated on static cross-sectional CT images without longitudinal follow-up data. Tracking nodule evolution over time is clinically important for distinguishing slow-growing indolent lesions from rapidly progressing cancers and for monitoring treatment response. Incorporating temporal information into future model versions would substantially enhance clinical utility.
Future Research Directions. Future work should expand datasets to include more diverse nodule types and scanner protocols, explore integration of attention mechanisms or transfer learning to further enhance feature extraction, employ advanced augmentation or weighted loss functions to address potential class imbalance in clinical data, and conduct prospective clinical trials to validate performance across diverse real-world settings before routine deployment.