Hematology, Transfusion and Cell Therapy (Apr 2024)

ARTIFICIAL INTELLIGENCE TO EVALUATE METABOLIC TUMOR BURDEN IN PRIMARY STAGING OF RECTAL CANCER WITH 18F-FDG PET/CT

  • Victor Cabral Costa Ribeiro Heringer,
  • Maria Carolina S. Mendes,
  • Barbara Juarez Amorim,
  • Allan Oliveira Santos,
  • Marina N. Silveira,
  • Cleide Silva,
  • Juliano S. Fonseca,
  • Mariana C.L. Lima,
  • Lorena P. Cunha,
  • Carlos Augusto R. Martinez,
  • Claudio Coy,
  • Jose Barreto C. Carvalheira,
  • Elba Cristina Sá de Camargo Etchebehere

Journal volume & issue
Vol. 46
pp. S3 – S4

Abstract

Read online

Introduction/Justification: The use of artificial intelligence using convolutional neural networks in clinical practice is recent. Thus, a growing need exists to validate software performance in different tasks in different diseases. Objectives: To evaluate the performance of artificial intelligence software to determine metabolic tumor burden in the primary staging of rectal cancer. Materials and Methods: A cross-sectional retrospective analysis was conducted on 51 histology-proven rectal cancer patients (35% females; mean age = 61 years) who underwent a staging 18F-FDG PET/CT. Whole-body metabolic tumor burden parameters (wbMTV and wbTLG) were quantified semi-automatically and through AI algorithm (Syngovia VB60; Siemens Healthineers Medical Solutions). The AI software's ability to correctly identify and classify the primary lesion, regional lymph nodes, and distant metastases was evaluated. In addition, the intraclass correlation coefficient (ICC) was applied to evaluate concordance between the AI-based software and the semiautomatic software in determining wbMTV and wbTLG. Values above 0.7 were considered to indicate substantial agreement. Resultados: The AI and semiautomatic tumor burden metrics correlated strongly for both wbMTV (ICC = 1.00; 95% CI = 0.94 - 0.99; P < 0.0000) and wbTLG (ICC = 1.00; 95% CI = 0.80 – 0.90; P < 0.0000). Additionally, the AI software's ability to correctly identify lesions compared to the documented staging was better for the identification of distant metastasis (78,57% of patients), mildly adequate to identify regional lymph nodes (50,00%) and had poor performance for identification of the primary lesion (5,76%). On the other hand, the time spent calculating these metrics was less by AI than by the semiautomatic method, especially in patients with advanced disease. Conclusion: The determination of whole-body metabolic tumor burden on 18F-FDG PET/CT with artificial intelligence software is challenging because of the physiologic bowel activity. However, deep learning may have the ability to overcome these challenges and may therefore improve the primary staging of rectal cancer.

Keywords