The Radiation Oncology Feed For radiation oncologists · physicists · RTTs · dosimetrists
technical-relevance deep-learning AI

Deep learning predicted distant metastasis from CT and dose in head and neck cancer

A model integrating planning CT, 3D dose and tumor contours achieved strong external prediction of distant metastasis but weaker local-recurrence performance.

RadOncToday · ·1 min read

KEY POINTS

  • The study used 237 patients with head and neck cancer from four institutions in the public Head-Neck-PET-CT dataset. Patients had received definitive IMRT or VMAT, with prescriptions varying by center from 50–70 Gy in 20–35 fractions. The dataset was divided into 131 training, 65 internal-validation and 41 external-validation patients.
  • Unlike conventional imaging-only prognostic models, the dual-task network simultaneously processed planning CT, the complete 3D dose distribution and GTV masks, with shared and endpoint-specific network components predicting distant metastasis and locoregional recurrence.
  • External validation was strongest for distant metastasis: concordance reached 0.92 (95% CI 0.86–0.98), with 2- and 3-year AUCs of 0.95 and 0.96, respectively. Internal-validation concordance was 0.84.
  • Locoregional-recurrence prediction was less convincing. External-validation concordance was 0.74 (95% CI 0.59–0.86), with 2- and 3-year AUCs of 0.81 and 0.84. The wide confidence intervals around sensitivity and specificity reflect the limited number of events in the small external cohort.
  • The deep model outperformed the authors' conventional radiomics-based Cox models. Decision-curve analysis also favored deep learning in the external cohort, particularly for distant-metastasis prediction, although this remains retrospective model evaluation rather than evidence that acting on predictions improves outcomes.
  • Grad-CAM suggested that the two endpoints relied on different spatial information: distant-metastasis predictions emphasized lymphatic/peritumoral regions, whereas locoregional-recurrence predictions concentrated more strongly within and around the tumor.
  • The Activation Volume Histogram analysis identified the GTV plus 3 mm receiving ≥65 Gy as especially discriminative for distant-metastasis risk and the GTV receiving ≥65 Gy for locoregional recurrence. These maps are hypothesis-generating: Grad-CAM has limited spatial resolution, and the study does not establish that modifying dose in these regions changes recurrence risk.

CLINICAL TAKEAWAY

The notable advance is not simply the high distant-metastasis score—it is that the model combines what the tumor looked like with where the radiation dose actually went and attempts to show what drove its prediction. The external distant-metastasis performance is impressive, but 41 external patients are far too few to justify risk-adapted treatment based on the model today.

SOURCE

International Journal of Radiation Oncology, Biology, Physics