The Classification Challenge CT-based lung nodule classification requires distinguishing malignant from benign lesions with high sensitivity and specificity. Single deep learning architectures often perform well on benchmark datasets but struggle with the subtle textural and morphological heterogeneity seen in real clinical CT scans, particularly for small or ground-glass nodules.
Hybrid Architecture Design Published in Scientific Reports in 2025, this study proposes the ILN-TL-DM model - an Improved LeNet integrated with Transfer Learning and a DeepMaxout classifier. The architecture combines a novel segmentation network (P-ResU-Net), multi-source feature extraction, and an ensemble classification approach to achieve state-of-the-art lung cancer classification on the Luna16 benchmark dataset.
Luna16 Benchmark The study used the Lung Nodule Analysis 2016 (Luna16) dataset comprising 888 CT scans with 2,336 annotated instances. Luna16 is a widely used benchmark for lung nodule detection and classification algorithms, enabling direct comparison with published methods. The comprehensive four-stage pipeline of ILN-TL-DM addresses limitations in existing approaches at each processing step.
Adaptive Gaussian Filtering Preprocessing Raw CT images are preprocessed with an adaptive Gaussian filter that reduces noise while preserving edge information at nodule boundaries. Unlike fixed-parameter Gaussian smoothing, the adaptive approach adjusts filter parameters based on local image statistics, reducing over-smoothing of small nodule details that are critical for malignancy classification.
P-ResU-Net Segmentation Architecture The segmentation stage uses a modified UNet architecture incorporating Residual Dual Channel Attention Blocks (RDCAB). These attention blocks allow the network to selectively focus on the most informative feature channels - emphasizing edge-defining and texture-encoding channels while suppressing background noise channels. This produces more accurate nodule boundary delineation than standard UNet.
Novel Activation Functions The segmentation network uses two non-standard activation functions: PReLU (Parametric Rectified Linear Unit), which learns the negative slope parameter during training to preserve gradient flow in deeper layers, and d-SiLU (dynamic Sigmoid-weighted Linear Unit), which provides smoother gradient profiles than ReLU, reducing dead neuron effects and improving convergence on complex lung texture patterns.
LGTrP and PHOG Texture Features Local Gray-level Transition Pattern (LGTrP) extracts 50-dimensional features encoding transitions between neighboring pixel intensity levels - capturing fine-grained texture transitions that reflect nodule internal heterogeneity. Pyramid Histogram of Oriented Gradients (PHOG) also provides 50-dimensional features encoding gradient orientation distributions at multiple spatial scales, capturing shape and edge properties across resolution levels.
Deep Transfer Learning Features Pre-trained AlexNet and ResNet50 convolutional neural networks - trained on ImageNet - were fine-tuned on the lung CT domain to extract 100-dimensional deep feature vectors. Transfer learning leverages the general visual feature representations learned from millions of natural images to initialize the network, reducing the amount of lung CT training data needed for strong performance.
Improved Entropy Feature A single improved entropy measurement captures overall statistical complexity of the nodule intensity distribution. Though just one value, this entropy feature quantitatively summarizes whether the nodule interior has homogeneous (benign characteristics) or heterogeneous (malignant characteristics) intensity distribution. The four feature streams are concatenated into a 201-dimensional combined feature vector fed to the classifier.
Improved LeNet With Transfer Learning (ILN-TL) The classification stage uses a modified LeNet architecture augmented with VGG-16 transfer learning components. VGG-16's deep convolutional feature extraction is integrated into the LeNet backbone, combining VGG-16's proven image representation capabilities with LeNet's computationally efficient classification architecture. This hybrid avoids the computational overhead of full VGG-16 inference.
DeepMaxout Classifier The DeepMaxout network uses maxout activation units - each unit outputs the maximum of several linear projections rather than a fixed nonlinear function. This makes the network highly flexible in approximating arbitrary decision boundaries, particularly valuable for the irregular and non-linear separating surfaces between benign and malignant nodule classes in high-dimensional feature space.
Soft Voting Ensemble ILN-TL and DeepMaxout are run in parallel on the same 201-dimensional feature vector, and their final class probability outputs are combined by soft voting - averaging the probability distributions from both classifiers. Soft voting consistently outperforms hard voting (majority vote) because it incorporates the confidence levels of both models, giving more weight to high-confidence predictions.
Top-Line Classification Metrics The ILN-TL-DM model achieved accuracy 0.962, specificity 0.955, negative predictive value (NPV) 0.964, and AUC 0.985 on the Luna16 test set. The high NPV of 0.964 is particularly important clinically: it means that when the model predicts a nodule is benign, it is correct 96.4% of the time, reducing unnecessary interventions for benign lesions.
Ablation Study Results Ablation studies removing individual feature streams showed that each component contributed positively: removing LGTrP features reduced accuracy by approximately 2%; removing deep features had the largest impact (4% accuracy drop); and removing entropy reduced AUC by 1-2%. This confirms that the complementary feature streams are all functionally important rather than redundant.
Comparison With Prior Methods ILN-TL-DM outperformed all compared methods on Luna16, including state-of-the-art 3D CNN approaches, other transfer learning models, and radiomic feature-based classifiers. The combination of hand-crafted texture features with deep transfer learning features in the 201-dimensional vector provided a richer representation than either approach alone.
Single Benchmark Dataset Limitation The study evaluates ILN-TL-DM exclusively on Luna16, which was collected at specific scanner settings and annotated by a defined expert team. Performance on heterogeneous real-world clinical CT data from diverse scanner types, acquisition protocols, and patient populations may differ from the reported Luna16 benchmark numbers.
Binary Classification Scope The current model classifies nodules as benign or malignant. Clinical practice increasingly requires subclassification by histological subtype (adenocarcinoma, squamous cell, small cell) and molecular marker prediction (EGFR mutation status) to guide targeted therapy selection. Extending the model to multi-class histological and molecular prediction is a natural next research direction.
Prospective Clinical Validation To transition from a benchmark study to a clinical tool, ILN-TL-DM requires prospective validation in a multi-institutional clinical trial where CT scans from real screening or diagnostic pathways are processed by the system and compared against histopathology gold standards. Only such a trial can establish the true clinical sensitivity and specificity needed for regulatory approval and clinical adoption.