Multi-task learning-based histologic subtype classification of non-small cell lung cancer.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Computational Pathology

Background:

• Accurate histologic subtype classification of lung cancer is crucial for effective treatment planning.
• Non-small cell lung cancer (NSCLC) subtypes, adenocarcinoma and squamous cell carcinoma, require distinct therapeutic strategies.
• Current classification methods can be labor-intensive and rely on precise manual annotations.

Purpose of the Study:

• To evaluate the efficacy of a novel multi-task learning (MTL) model for histologic subtype classification of NSCLC using computed tomography (CT) images.
• To assess if MTL can improve classification accuracy compared to single-task approaches and radiomics methods.
• To determine if the MTL model can reduce the reliance on physician's precise labeling of tumor areas.

Main Methods:

• A novel MTL model was developed, incorporating a histologic subtype classification branch and a staging branch that share feature extraction layers.
• The model was trained simultaneously on both tasks.
• A dataset of 402 NSCLC cases from The Cancer Imaging Archive (TCIA) was utilized, split into training, internal test, and external test sets.

Main Results:

• The MTL model achieved an Area Under the Curve (AUC) of 0.843 on the internal test set and 0.732 on the external test set.
• The multi-task network demonstrated higher accuracy and specificity compared to single-task networks.
• The proposed model achieved superior performance over traditional radiomics methods.

Conclusions:

• Multi-task learning, by sharing network layers, significantly improves the accuracy of histologic subtype classification for NSCLC.
• The developed MTL model alleviates the need for precise manual labeling of lesion regions by physicians.
• This approach has the potential to reduce the manual workload for physicians, streamlining the diagnostic process.