Deep Learning Approach to Automatize TMTV Calculations Regardless of Segmentation Methodology for Major FDG-Avid...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Total metabolic tumor volume (TMTV) is an emerging prognostic factor in lymphoma.
• Manual segmentation of TMTV from 18FDG-PET/CT scans is time-consuming.
• Deep learning offers potential for automating TMTV calculation.

Purpose of the Study:

• To develop and validate a deep learning model for automated TMTV segmentation in lymphomas.
• To assess the accuracy and efficiency of the automated model compared to manual segmentation.
• To establish a generic deep learning model for TMTV computation.

Main Methods:

• A 3D V-NET convolutional neural network (CNN) was trained using 1218 baseline 18FDG-PET/CT scans.
• Ground truth segmentation was derived from manual regions of interest using various thresholds (TMTVprob).
• The model was trained with soft dice loss and validated on separate datasets (407 test, 405 validation).

Main Results:

• The deep learning model achieved mean Dice scores of 0.84 (training), 0.84 (validation), and 0.76 (test).
• Median Dice scores varied by cutoff method (0.77 for 41%, 0.70 for 2.5 SUV, 0.90 for 4 SUV).
• Spearman correlations between manual and predicted TMTV were high (0.92-0.98), with minimal volume differences (5-147 mL).

Conclusions:

• A generic deep learning model can accurately and efficiently compute TMTV in lymphomas.
• Automation of TMTV calculation using CNNs significantly reduces processing time.
• This automated approach can aid in assessing prognostic factors in lymphoma patients.