💻 computer science

Lung Cancer Classification from CT Images: A Comparative Study of SVM, Modified CNN, and EfficientNet–SVM Hybrid Models

This paper presents a comprehensive computer-aided diagnosis framework for lung cancer classification on the IQ-OTH/NCCD CT dataset, demonstrating that a modified CNN architecture enriched with batch normalization outperforms both traditional SVM-based approaches and a hybrid EfficientNet–SVM model by achieving a test accuracy of 99.41%.

Original authors: Nuran Binte Habib, Shahida Akther, Rabiul Islam

Published 2026-11-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nuran Binte Habib, Shahida Akther, Rabiul Islam

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Lung cancer remains the leading cause of cancer death worldwide, claiming millions of lives annually. The difference between survival and tragedy often comes down to timing. When the disease is caught early, while it is still a small, localized spot, the chance of surviving five years is high. Once it spreads, that chance drops precipitously. Doctors rely heavily on computed tomography, or CT scans, to find these early signs. A CT scan creates a detailed, three-dimensional map of the lungs by taking hundreds of thin cross-sectional images. However, reading these scans is a demanding task for human radiologists. A single patient's scan can contain hundreds of slices, and the early nodules that signal cancer can be subtle, easily missed, or confused with harmless spots. This cognitive load and the potential for human error have driven researchers to develop computer systems that can assist in spotting these dangerous patterns.

In this context, a team of researchers from Bangladesh and the United States set out to build and test a new kind of computer-aided diagnosis system. They focused on a specific challenge: teaching a computer to look at CT scan slices and correctly sort them into three categories—normal lung tissue, benign (harmless) nodules, and malignant (cancerous) nodules. The researchers worked with a public collection of over a thousand CT slices from 110 patients. A major hurdle in this work is that the data is unbalanced; there are far more images of cancerous lungs than of harmless ones. If a computer is trained on such uneven data, it tends to ignore the rare cases, which is exactly the opposite of what doctors need. To solve this, the team did not rely on a single method. Instead, they built a comprehensive framework that tested four different approaches side-by-side to see which one worked best.

The first approach the team explored was a traditional method that relies on human-designed rules. They took the raw CT images and cleaned them up using a series of steps: smoothing out the grainy noise, brightening the contrast to make features stand out, and tracing the edges of the lung tissue to isolate it from the rest of the body. From these cleaned images, they extracted specific numerical descriptions of the texture, shape, and brightness of the spots. They then fed these numbers into a mathematical classifier known as a Support Vector Machine. To handle the imbalance in the data, they tested six different strategies to artificially balance the number of examples the computer saw, such as duplicating rare cases or creating new, synthetic examples that look like the rare ones. They also tested three different ways for the computer to measure the distance between data points. The most successful version of this traditional approach used a specific strategy to balance the data and a non-linear way of measuring distances, achieving a balanced accuracy of nearly 99 percent.

The second approach abandoned the human-designed rules entirely and let the computer learn the patterns on its own using a deep learning model called a Convolutional Neural Network, or CNN. The team started with a basic version of this network, which is essentially a stack of layers that scan an image to find simple edges, then shapes, and finally complex structures. This basic model performed very well, correctly identifying the lung condition in nearly 99 percent of the test cases. However, the researchers made a specific modification to improve it further. They added a step called batch normalization, which acts like a stabilizer for the learning process, helping the computer adjust its internal settings more efficiently as it trains. This single change pushed the accuracy even higher, to 99.41 percent, making it the most successful model in their study.

The third approach was a hybrid, combining the strengths of the previous two. The researchers used a powerful, pre-trained deep learning model called EfficientNet to act as a feature extractor. This model, which had already learned to recognize objects in millions of everyday photos, was used to convert each CT slice into a compact mathematical summary. Instead of letting the deep learning model make the final decision, they passed this summary to a Support Vector Machine to make the final call. This separation of duties allowed the system to leverage the deep learning model's ability to see complex patterns while using the robust decision-making of the traditional classifier. This hybrid model achieved an accuracy of 98 percent, proving that combining different types of intelligence can yield strong results without needing the most massive, complex systems available.

The researchers compared their findings against other recent studies that used the same dataset. While some other teams achieved similar high scores, they often relied on more complex techniques, such as generating fake images to train the system or using massive, heavy-duty neural networks. The team's modified CNN achieved the highest accuracy of all, 99.41 percent, without needing these extra, computationally expensive steps. They also noted that their system was particularly good at identifying malignant cases, which is the most critical task for saving lives. The study concludes that a carefully tuned, moderately sized computer model can rival the performance of much larger, more complex systems. By systematically testing different ways to handle data imbalance and different architectural tweaks, the researchers demonstrated that high performance in medical imaging does not always require the most complicated tools, but rather a thoughtful combination of established methods.

Technical Summary: Lung Cancer Classification from CT Images

Problem Statement
Lung cancer remains the leading cause of cancer mortality globally, with early detection being critical for improving the five-year survival rate. While Computed Tomography (CT) is the modality of choice for screening, the volumetric nature of CT scans (typically 80–200 slices per case) imposes a significant cognitive load on radiologists, leading to potential oversight of subtle early-stage nodules. Furthermore, existing automated studies on the IQ-OTH/NCCD dataset often suffer from methodological gaps: they frequently evaluate deep learning or classical machine learning models in isolation rather than comparatively, lack systematic ablation studies on class-imbalance handling strategies, and rely on limited evaluation metrics that may not fully capture performance on skewed, multi-class distributions.

Methodology
The authors propose a unified Computer-Aided Diagnosis (CAD) framework evaluated on the IQ-OTH/NCCD dataset, comprising 1,097 CT slices from 110 patients across three classes: Normal (416), Benign (120), and Malignant (561). The framework integrates six sequential stages:

Image Preprocessing: A deterministic pipeline applies Gaussian smoothing ( $\sigma=1$ ), histogram equalization, Canny edge detection (thresholds 50/150), and morphological closing to suppress noise and enhance lesion conspicuity.
Lung Segmentation: Otsu's thresholding is used for initial segmentation, followed by morphological cleaning, small-object removal, and hole filling to generate a binary lung mask.
Feature Extraction (for Classical Models): A handcrafted feature vector is constructed from the segmented region, combining four categories:
- Statistical: Mean, standard deviation, variance, median, min/max, percentiles, skewness, and kurtosis.
- Texture: Local Binary Patterns (LBP) including rotation-invariant variants.
- Gradient: First-order spatial derivatives (magnitude and orientation) via Sobel kernels.
- Shape: Area, perimeter, compactness, aspect ratio, extent, and solidity.
Class Imbalance Handling: To address the severe imbalance (approx. 4.67:1 ratio between malignant and benign classes), six strategies are systematically evaluated: SMOTE, ADASYN, Borderline-SMOTE, Random Over-Sampling (ROS), SMOTE–Tomek, and cost-sensitive class weighting.
Model Architectures: Three distinct classification paradigms are implemented and compared:
- SVM: A Support Vector Machine benchmarked across a $6 \times 3$ ablation grid, testing six imbalance strategies against three kernels (Linear, RBF, Polynomial).
- Baseline CNN: A custom 3-block convolutional network (32, 64, 128 filters) with max-pooling, dense layers, and dropout, trained without batch normalization.
- Modified CNN: An enhanced version of the baseline CNN incorporating Batch Normalization after every convolutional operation, prior to ReLU activation, to accelerate convergence and reduce internal covariate shift.
- Hybrid EfficientNet–SVM: A model using a pre-trained EfficientNet-B0 backbone (frozen) to extract 1,280-dimensional embeddings, which are then fed into a kernel SVM. This approach decouples deep representation learning from maximum-margin decision boundary fitting.
Evaluation: Beyond standard accuracy, the study employs imbalance-aware metrics including Balanced Accuracy, Macro-averaged AUC-ROC, Matthews Correlation Coefficient (MCC), and Cohen's $\kappa$ .

Key Contributions
The paper claims five primary contributions:

Unified Pipeline: A reproducible, lightweight image-processing and feature-extraction pipeline combining deterministic preprocessing with a four-category handcrafted feature set.
Systematic Ablation: The most comprehensive comparison to date on the IQ-OTH/NCCD dataset, evaluating a $6 \times 3$ grid of imbalance-handling strategies and SVM kernels.
Modified CNN Performance: Demonstration that a simple architectural modification (adding Batch Normalization) to a baseline CNN yields 99.41% test accuracy without requiring GAN-based augmentation or complex attention mechanisms.
Hybrid Model Validation: Presentation of an EfficientNet-B0–SVM hybrid that achieves 98.00% accuracy with perfect recall on malignant cases, illustrating the efficacy of decoupling feature extraction from classification.
Rigorous Evaluation: Inclusion of MCC and Cohen's $\kappa$ to provide a more faithful characterization of multi-class performance under skewed distributions.

Results
The experimental results on the held-out test set (15% of data) are as follows:

Modified CNN: Achieved the highest overall test accuracy of 99.41% with a macro F1-score of 0.99. It demonstrated perfect precision for benign and malignant cases and near-perfect performance on normal cases.
SVM (ROS + RBF): The best-performing classical model achieved a balanced accuracy of 98.81% and a macro F1-score of 98.81%. The RBF kernel consistently outperformed Linear and Polynomial kernels across all imbalance strategies.
Hybrid EfficientNet–SVM: Attained 98.00% test accuracy with a macro F1-score of 0.957. Notably, it achieved perfect recall (1.00) for malignant cases on the test set.
Baseline CNN: Achieved 98.79% accuracy, serving as a reference point that showed the Modified CNN's improvement was attributable to Batch Normalization.

Significance and Claims
The authors assert that their work demonstrates competitive, state-of-the-art performance on the IQ-OTH/NCCD dataset without resorting to computationally expensive GAN-based augmentation or large transformer backbones. The study highlights that:

Simplicity vs. Complexity: A Modified CNN with Batch Normalization can outperform more complex architectures (e.g., Lung-EffNet with GANs) while maintaining a simpler topology.
Imbalance Strategy: Explicit oversampling (specifically Random Over-Sampling and ADASYN) combined with RBF kernels is superior to cost-sensitive weighting alone for this specific dataset's imbalance profile.
Hybrid Viability: Hybrid models can effectively leverage pre-trained deep features for representation while utilizing the robust decision boundaries of SVMs, offering a balance between deep learning power and classical interpretability.

The paper concludes that while the dataset size (110 patients) is a limitation, the proposed framework offers a robust, reproducible baseline for lung cancer classification, with future work suggested to include multi-center validation and the integration of spatio-temporal architectures.

Technical Summary: Lung Cancer Classification from CT Images

More like this