Histological Subtypes Classification of Lung Cancers on CT Images Using 3D Deep Learning and Radiomics

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiomics

Background:

• Accurate histological subtyping of lung cancer is crucial for effective clinical treatment decisions.
• Distinguishing between lung adenocarcinoma (ADC), squamous cell carcinoma (SCC), and small cell lung cancer (SCLC) can be challenging.
• Non-invasive methods for predicting lung cancer subtypes are highly desirable.

Purpose of the Study:

• To develop and compare the performance of 3D deep learning and radiomics models for the automatic classification of lung cancer histological subtypes.
• To assess the ability of these models to distinguish between ADC, SCC, and SCLC using Computed Tomography (CT) images.

Main Methods:

• A retrospective study included 920 patients with histopathologically confirmed lung cancer (554 ADC, 175 SCC, 191 SCLC).
• Two classification models were designed: ProNet (3D deep learning) and com_radNet (radiomics).
• Models were trained, validated, and tested on 70%, 15%, and 15% of the dataset, respectively.

Main Results:

• The ProNet model achieved weighted average F1-scores of 90.0% (ADC), 72.4% (SCC), and 83.7% (SCLC), with an overall weighted average F1-score of 73.2%.
• The com_radNet model achieved F1-scores of 83.1% (ADC), 75.4% (SCC), and 85.1% (SCLC), with an overall weighted average F1-score of 72.2%.
• ProNet demonstrated an AUC of 0.840 and accuracy of 71.6%, while com_radNet had an AUC of 0.789 and accuracy of 74.7%.

Conclusions:

• Both ProNet and com_radNet models demonstrate high performance in distinguishing between ADC, SCC, and SCLC.
• These AI-driven approaches show potential as non-invasive tools for predicting lung cancer histological subtypes.
• Further validation may establish these methods as valuable aids in clinical decision-making for lung cancer treatment.