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

This study investigates whether multi-label learning improves PET/CT-based radiogenomic prediction of EGFR, TP53, and KRAS mutations in non-small cell lung cancer (NSCLC) compared with conventional single-gene classification. The authors state: To the best of our knowledge, this is among the first studies to systematically investigate multi-label learning for PET/CT radiogenomic mutation prediction in NSCLC.

The study was conducted on a novel UK-based radiogenomics cohort, the MIRAI dataset, which contains retrospectively collected, real-world clinical data from a single hospital between November 2014 and October 2024. The final cohort included 263 patients with a confirmed diagnosis of NSCLC and 293 unique whole-body 18F-FDG PET/CT examinations. Of these, 108 had corresponding gene mutation labels available and were included in the final gene mutation prediction dataset. Tumours were either manually segmented or automatically segmented using a pretrained nnU-Net model, which achieved a mean Dice coefficient of 0.78 on the test set.

A ResNet-18 pretrained on ImageNet-1K was used as the backbone, with the final layer replaced with either one output (single-gene) or two outputs (multi-label). Inputs consisted of axial PET and CT slices cropped to the largest cross-sectional area of the primary tumour, concatenated with the segmentation mask. Models were trained with binary cross-entropy loss with positive weights to address class imbalance, for a maximum of 30 epochs with early stopping, and evaluated using AUC, precision, recall, and F1-score across five random seeds.

The results show that the effectiveness of multi-label learning depends on the gene pair. For the KRAS and TP53 pair, joint learning improved both tasks: the AUC increasing from 0.575 to 0.641 for KRAS and from 0.689 to 0.712 for TP53 compared with the corresponding single-gene models. For the EGFR/KRAS pair, joint learning improved EGFR prediction but reduced KRAS performance, with the authors noting negative transfer may occur for KRAS. For the EGFR/TP53 pair, joint learning degraded performance for both tasks, with single-gene models achieving higher AUCs. The authors note that no EGFR-positive/TP53-positive cases remained in the modelling dataset after preprocessing due to rare label combinations being excluded.

The authors conclude: "the effectiveness of multi-label learning for PET/CT-based radiogenomic mutation prediction depends on the specific combination of gene mutations being modelled. Specifically, joint learning improved the prediction of KRAS and TP53 mutations but did not improve performance for the TP53/EGFR pair. These findings suggest that multi-label learning should not be applied uniformly across all radiogenomic prediction tasks. Instead, mutation-specific modelling strategies may be preferable, with joint learning adopted only when it provides measurable performance gains."

The study acknowledges several limitations: the dataset exhibits class imbalance, the study was conducted using data from a single centre without external validation, tissue biopsy samples were obtained from a single region of each tumour, and this work was limited to single 2D image analysis and did not explore the use of 3D models. Future work will address class imbalance, validate on external multi-centre cohorts, and explore 3D models.

Improvements to AI Systems:

1. Adaptive Multi-Label Learning with Negative Transfer Detection

• Implement a dynamic training framework that automatically assesses pairwise label dependencies before joint learning. Use a meta-learning or gradient-based approach to quantify task similarity (e.g., cosine similarity of gradient directions) and decide whether to share a backbone or use separate heads per gene.

• Improved system capability: Automatically avoids negative transfer for pairs like EGFR/TP53 by switching to independent single-task heads when joint training degrades performance, while still benefiting from joint learning for KRAS/TP53.

2. Label Co-occurrence-Aware Data Augmentation and Sampling

• Incorporate a generative model (e.g., conditional VAE or diffusion) to synthesize synthetic PET/CT slices for rare label combinations (e.g., EGFR+/TP53+), conditioned on the segmentation mask and mutation labels. This addresses the exclusion of rare combinations during preprocessing.

• Improved system capability: Enables training on all possible mutation combinations, improving generalization to underrepresented phenotypes and reducing bias from dropped samples.

3. Uncertainty-Aware Multi-Label Prediction with Confidence Calibration

• Add a probabilistic output layer (e.g., Monte Carlo dropout or deep ensembles) to produce per-gene prediction intervals. Use these uncertainties to trigger a fallback to single-gene models when joint predictions have high variance, or to flag low-confidence cases for manual review.

• Improved system capability: Provides clinicians with reliability scores, reducing false positives/negatives in critical mutation screening, and automatically escalates ambiguous cases to dedicated single-gene classifiers.

4. 3D Volumetric Multi-Label Radiogenomic Transformer

• Replace the 2D ResNet-18 with a 3D vision transformer (e.g., Swin-UNETR) that processes full PET/CT volumes, using the segmentation mask as a token-level attention mask. Train with a multi-task loss that includes a label-correlation matrix as a regularizer.

• Improved system capability: Captures spatial heterogeneity of tumours across slices, improving mutation prediction accuracy beyond single-slice analysis, and learns joint feature representations that respect gene co-occurrence patterns.

5. Cross-Centre Domain Adaptation with Federated Multi-Label Learning

• Develop a federated learning framework where multiple hospitals train local multi-label models on their own PET/CT data, sharing only model gradients. Use adversarial domain adaptation to align feature distributions across centres, with a meta-learner that optimizes the trade-off between joint and single-task performance per centre.

• Improved system capability: Enables robust, generalizable mutation prediction across diverse clinical settings without sharing patient data, and automatically adapts to centre-specific imaging protocols while preserving multi-label benefits where they exist.

6. Explainable Multi-Label Decision Support with Counterfactual Generation

• Integrate an explainability module (e.g., Grad-CAM or SHAP) that highlights which tumour subregions drive each gene prediction. Additionally, generate counterfactual PET/CT slices (e.g., via adversarial perturbation) to show what changes would flip a mutation prediction, aiding clinician trust and hypothesis generation.

• Improved system capability: Provides actionable insights into why a specific mutation is predicted, supports biopsy guidance by identifying informative tumour regions, and helps radiologists verify model reasoning in real time.

Lung cancer remains one of the leading causes of cancer- related mortality worldwide. Although targeted therapies have improved outcomes for patients with non-small cell lung cancer (NSCLC), they rely on mutation profiling through tissue biopsy, an invasive procedure with several limitations. This study investigates PET/CT-based radio- genomic prediction of epidermal growth factor receptor (EGFR), tumour protein 53 (TP53), and Kirsten rat sarcoma viral oncogene (KRAS) mutations using deep learning. We further evaluate whether pairwise multi-label learning improves mutation prediction compared with conventional single-gene classification. To the best of our knowledge, this is among the first studies to systematically investigate multi-label learning for PET/CT radiogenomic mutation prediction in NSCLC. Experiments were conducted on a novel UK-based radiogenomics cohort. Joint pre- diction of KRAS and TP53 improved AUC from 0.58 to 0.64 for KRAS and from 0.69 to 0.71 for TP53. For the EGFR/KRAS pair, only EGFR benefited from joint learning, while no improvement was observed for the EGFR/TP53 pair. These findings demonstrate that the effectiveness of multi-label learning depends on the specific combination of gene mutations being modelled, suggesting that mutation-specific modelling strategies may be preferable for PET/CT radiogenomic prediction.