Optimized Wavelet Scatter Method for Accurate Classification and Segmentation of Lung Nodules

Abstract

       Lung cancer is a significant global health concern due to its high mortality rate. This is primarily due to the difficulty of identifying malignant growths in early-stage computed tomography (CT) images. The thing is, it's often hard to tell the difference between a benign nodule and a malignant one. This makes it tricky to figure out what's going on with the patient. That's where these computer-aided diagnosis (CAD) approaches come in. This study introduces a lightweight and interpretable framework for comprehensively analyzing lung nodules, including their detection, classification, and segmentation. This approach uses the wavelet scattering transform (WST) to extract stable, deformation-invariant features. These features are then evaluated using three traditional classifiers: The Support Vector Machine (SVM), the Random Forest (RF), and the Decision Tree (DT). The methodology follows a structured pipeline. It begins with image enhancement steps, such as resizing, denoising, and gamma-based contrast optimization. Then, it moves on to WST-based feature representation, classification, and segmentation using Otsu’s thresholding technique. Validation experiments were conducted using the SPIE-AAPM Lung CT Challenge dataset in conjunction with the LIDC-IDRI benchmark. This dataset comprises 11,114 annotated CT slices, of which 8,286 are malignant and 2,828 are benign. The results demonstrate the proposed pipeline's effectiveness in delivering accurate and reliable performance. This shows its potential as a supportive tool for early diagnosis and clinical decision-making. Patient-wise separation and stratified 10-fold cross-validation were applied to ensure unbiased evaluation and prevent data leakage. Results show that SVM achieved perfect performance across all metrics (accuracy, precision, recall, and F1-score), while RF and DT yielded 99.33% and 98.18%, respectively. For segmentation, the model reached an Intersection over Union (IoU) of 0.845 and a Dice Similarity Coefficient (DSC) of 0.912. Although the SVM classifier achieved ideal scores, measures such as patient-level grouping, class weighting, and cross-validation were implemented to reduce risks of overfitting. Overall, the integration of WST with classical classifiers provides a lightweight and explainable CAD framework that has strong potential to support early lung cancer detection and clinical decision-making.