Predicting malignant nodules by fusing deep features with classical radiomics features.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiology

Background:

• Lung cancer presents high incidence and mortality rates.
• Early detection via low-dose computed tomography (CT) is crucial.
• Radiomics and deep learning (convolutional neural networks - CNNs) show promise for lung cancer diagnosis and prognosis.

Purpose of the Study:

• To differentiate lung cancer nodules from positive controls using a transfer learning approach.
• To explore the efficacy of deep features extracted by pretrained CNNs and combined with radiomics features.
• To develop and evaluate CNN models for lung cancer detection using National Lung Screening Trial (NLST) data.

Main Methods:

• Utilized subsets of participants from the National Lung Screening Trial (NLST).
• Employed transfer learning with three different pretrained CNNs for deep feature extraction.
• Experimented with various classifiers, deep features from different color channels, and combinations of deep and radiomics features.
• Designed and trained a CNN on augmented NLST data.

Main Results:

• Feature combinations yielded the best classification accuracy of 76.79%.
• A CNN trained on augmented NLST data achieved an area under the receiver operating characteristic curve of 0.87.
• Deep features from pretrained CNNs and classical radiomics were effective in analysis.

Conclusions:

• Transfer learning and deep learning approaches, particularly when combining feature types, enhance the accuracy of lung cancer nodule detection.
• CNNs trained on augmented medical imaging data show significant potential for improving diagnostic performance in oncology.
• The study demonstrates a viable strategy for leveraging limited medical datasets through advanced AI techniques.