Development of a Multimodal AI Model to Predict the Efficacy of Photodynamic Therapy for Local Residual or Recurrent Lesions Following Chemoradiotherapy or Radiotherapy for Esophageal Cancer.

OBJECTIVES

Approximately 12%-40% of patients with esophageal cancer develop local residual or recurrent lesions after chemoradiotherapy (CRT) or radiotherapy (RT). Photodynamic therapy (PDT) is a minimally invasive, organ-preserving alternative; however, therapeutic response is not guaranteed. We developed an artificial intelligence (AI) model to predict PDT efficacy.

METHODS

We retrospectively analyzed 177 patients who underwent PDT after CRT/RT at the National Cancer Center Hospital East between August 2006 and December 2024. Clinical data and pre-PDT endoscopic images obtained within 1 month were collected. Local image features extracted via oriented FAST and rotated BRIEF were integrated with clinical information to build a multimodal support vector machine for binary classification. Model performance was evaluated on an independent test set and compared with that of three endoscopists.

RESULTS

The AI model was trained on 145 cases, achieving an accuracy, recall, precision, and F1 score, and area under the receiver operating characteristic curve of 0.813, 0.833, 0.833, 0.833, and 0.897, respectively in the test cohort of 32 cases. The AI model showed numerically higher accuracy, recall, precision, and F1 score than the endoscopists (0.531, 0.636, 0.389, and 0.483, respectively).

CONCLUSIONS

A multimodal AI model integrating clinical data and endoscopic images can accurately predict PDT efficacy after CRT or RT. This approach may help optimize treatment strategies and individualize patient management.