PET/CT Radiogenomic Mutation Prediction in Non-Small Cell Lung Cancer Using Multi-Label Learning

Technical Summary: PET/CT Radiogenomic Mutation Prediction in Non-Small Cell Lung Cancer Using Multi-Label Learning

Problem Statement
Non-small cell lung cancer (NSCLC) treatment increasingly relies on molecular profiling to guide targeted therapies, such as EGFR tyrosine kinase inhibitors. However, the current clinical standard for identifying mutations (EGFR, TP53, KRAS) is invasive tissue biopsy, which suffers from limitations including single temporal snapshots, limited spatial sampling that may miss genetic heterogeneity, and challenges in accessing certain anatomical sites. While radiogenomics offers a non-invasive alternative by inferring genomic characteristics from medical imaging, most existing studies focus on predicting the status of individual genes using CT alone. There is a lack of systematic investigation into whether jointly predicting multiple mutations using PET/CT data via multi-label learning can improve predictive performance compared to independent single-gene classification.

Methodology
The study utilized the MIRAI dataset, a real-world UK-based cohort of 263 NSCLC patients with 293 whole-body 18F-FDG PET/CT examinations. Genetic testing was performed within a ±6-month window of the imaging acquisition.

• Data Preprocessing:

  • Segmentation: A total of 130 PET/CT volumes were manually segmented using ITK-SNAP. The remaining volumes were segmented using a pretrained nnU-Net model (fine-tuned on 110 manual cases and evaluated on 20). Lymph nodes and metastatic lesions were excluded, retaining only the primary tumor mask.
  • Image Preparation: PET and CT volumes were resampled to 3.27 mm isotropic spacing. For each patient, axial slices corresponding to the largest cross-sectional area of the tumor were extracted, cropped via bounding box, and resized to 224 × 224 pixels.
  • Input Construction: The PET slice, CT slice, and segmentation mask were concatenated along the channel dimension. CT intensities were clipped to the 0.5th–99.5th percentiles and Z-score normalized; PET slices were globally Z-score normalized.

• Model Architecture:

  • A ResNet-18 backbone (pretrained on ImageNet-1K) was used for all models.
  • Single-Gene Models: The final fully connected layer was replaced with a single output for binary classification.
  • Multi-Label Models: The final layer was replaced with two outputs to predict two gene mutations simultaneously.
  • Training: All network parameters were updated (no freezing of the backbone). Models were trained using binary cross-entropy loss with logits and positive class weighting to address imbalance. Data augmentation included random horizontal flips, affine transformations, and Gaussian noise.
  • Experimental Design: Patients were split into training, validation, and test sets (70:15:15) using stratification based on target labels. Repeated scans from the same patient were treated as independent samples. To ensure fair comparison, single-gene models were trained on the exact same patient splits as their corresponding multi-label counterparts.

Key Contributions

1. Systematic Comparison: This is among the first studies to systematically compare single-gene and pairwise multi-label deep learning models for PET/CT-based radiogenomic mutation prediction in NSCLC.
2. Multi-Modal Approach: The study investigates multi-gene classification using both PET and CT imaging, a relatively underexplored area compared to CT-only studies.
3. Novel Cohort: The evaluation was conducted on a novel real-world UK PET/CT radiogenomics cohort, addressing the underrepresentation of this specific patient population in existing literature.

Results
The study evaluated the prediction of EGFR, KRAS, and TP53 mutations across three pairwise combinations:

• KRAS and TP53 Pair: Joint multi-label learning improved performance for both tasks. The Area Under the Curve (AUC) increased from 0.575 to 0.641 for KRAS and from 0.689 to 0.712 for TP53 compared to single-gene models. This suggests positive transfer between these tasks.
• EGFR and KRAS Pair: Joint learning improved EGFR prediction (AUC 0.718 to 0.746) but reduced KRAS performance (AUC 0.632 to 0.557). The authors suggest this may be due to negative transfer or task imbalance, where the larger positive class of EGFR influences optimization more heavily, or because EGFR imaging features are more discriminative than KRAS features.
• EGFR and TP53 Pair: Joint learning degraded performance for both tasks compared to single-gene models (EGFR AUC: 0.808 vs. 0.783; TP53 AUC: 0.514 vs. 0.491). This indicates limited complementary information between these specific mutations, potentially leading to gradient conflict during optimization. Notably, no EGFR-positive/TP53-positive cases remained in the dataset after preprocessing, limiting the analysis of co-mutation.

Significance and Claims
The paper claims that the effectiveness of multi-label learning for PET/CT radiogenomic mutation prediction is not uniform; it depends heavily on the specific combination of gene mutations being modeled.

• Modest Conclusion: The authors conclude that multi-label learning should not be adopted as a universal strategy for all radiogenomic tasks. Instead, mutation-specific modeling strategies are preferable.
• Recommendation: Joint learning should be applied only when it provides measurable performance gains, as demonstrated in the KRAS/TP53 pair, while independent modeling may remain superior for other combinations like EGFR/TP53.
• Limitations: The study acknowledges limitations including class imbalance (reflecting real-world prevalence but affecting model training), lack of external validation (single-center data), potential sampling bias from single-region biopsies, and the restriction to 2D image analysis rather than 3D models.