Acute side effect profiles in oropharyngeal cancer; Deep learning-based versus Manual IMPT Planning

Objective
Recent advances in deep learning (DL)-based dose optimization have improved efficiency. However, their impact on clinical outcomes, particularly common side effects, remains unclear. This study investigated whether DL-based IMPT plans resulted in comparable or reduced acute side effects compared with manually created IMPT plans in patients with oropharyngeal cancer (OPC).

Methods
DL-based plans were generated in RayStation in three steps: 1) a U-net DL model predicted dose distributions; 2) these predicted dose distributions were converted into clinically deliverable plans using a mimicking algorithm; and 3) dosimetrists’ adjustments when needed (within 2 hours). Clinical outcomes of 88 OPC patients treated with bilateral DL-based plans (prescribed dose: 70 Gy and 54.25 Gy) were compared to those of 132 OPC patients recently treated with manually planned IMPT. Acute side effects were prospectively assessed using questionnaires (EORTC HN35) and CTCAE v4.0 criteria at baseline, weekly during treatment, and at 12 weeks and 6 months following treatment. Primary endpoints included: 1) Physician-rated dysphagia (Grade ≥ 2 and Grade ≥ 3), and 2) Patient-rated xerostomia (moderate–severe and severe). Differences in probability of side effects in the final week of treatment between manual IMPT and DL-based IMPT were estimated using logistic regression with adjustment for baseline complaints. The secondary endpoint included the total toxicity burden, defined as the sum of 15 acute toxicities (e.g. dysphagia, mucosal reactions, dry mouth, dermatitis, tube feeding) weighted by severity and duration (from baseline to 6 months after treatment).

Results
The two cohorts had comparable mean organ doses (Figure 1), except for a difference in Pharyngeal Constrictor Muscle (PCM) inferior mean dose. In the final week of treatment, Grade ≥ 2 dysphagia was reported in 76.5% of patients receiving manual IMPT versus 78.4% among those treated with DL-based IMPT (p = 0.876), and Grade ≥ 3 dysphagia reported in 46.2% and 48.9% (p = 0.862), respectively (Fig. 2). Moderate to severe xerostomia was reported in 42.4% (manual IMPT) versus 43.2% (DL-based IMPT, p=0.972), and severe xerostomia was reported in 10.6% and 12.5% (p=0.653) of patients, respectively. The average total toxicity burden was 34 points for manual IMPT and 38 points for DL-based IMPT.

Conclusion
This is the first prospective study to show that DL-based IMPT planning yields acute side-effect profiles comparable to those observed among OPC patients treated with manually planned IMPT. These findings indicate that DL-based IMPT planning can uphold clinical standards while offering substantial gains in planning efficiency.