Application of Fractal Radiomics and Machine Learning for Differentiation of Non-Small Cell Lung Cancer Subtypes on...

Area of Science:

• Radiology and Medical Imaging
• Oncology
• Machine Learning in Medicine

Background:

• Non-small cell lung cancer (NSCLC) is the most common lung malignancy, with adenocarcinoma (ADC) and squamous cell carcinoma (SCC) requiring different treatments.
• Accurate differentiation of NSCLC subtypes using diagnostic imaging is clinically significant but challenging.
• Radiomic features, including first-order statistics (FOS), second-order statistics (SOS), and fractal dimension texture analysis (FDTA), extracted from magnetic resonance (MR) images, offer quantitative assessment potential.

Purpose of the Study:

• To assess if integrating FDTA features with FOS and SOS texture features in MR image analysis enhances machine learning classification of NSCLC into ADC and SCC subtypes.
• To evaluate the diagnostic performance of combined radiomic features for NSCLC subtype differentiation.

Main Methods:

• A dataset of 274 MR images (122 ADC, 152 SCC) was analyzed.
• 93 texture features were extracted from segmented MR images.
• Random forest was used for feature selection from FOS/SOS and combined FOS/SOS/FDTA datasets, followed by k-nearest neighbors (kNN) classification.

Main Results:

• The combined FOS/SOS/FDTA dataset yielded the highest classification performance.
• Feature selection identified 37 informative texture features from the combined dataset.
• The best model achieved an accuracy of 0.78, precision of 0.81, and AUC of 0.89.

Conclusions:

• Incorporating fractal descriptors significantly enhances texture-based classification of lung MR images for NSCLC subtypes.
• This approach improves the differentiation between adenocarcinoma and squamous cell carcinoma.
• The findings suggest potential for improved quantitative assessment and personalized treatment strategies in NSCLC.