Author of 1 Presentation
SS 5.4 - Deep learning for fully automated segmentation of rectal tumours on MRI in a multicenter setting
Abstract
Purpose
To explore the performance of deep learning to perform fully automated segmentation of rectal tumors on baseline MRI using a multicenter dataset (with variations in acquisition protocols and scan quality reflecting daily practice).
Material and methods
Baseline staging MRIs of 355 patients (from 6 centers) were analyzed. Overall scan quality was assessed using a 5-point Likert score (0=poor) to 4=excellent). Data were randomly split into train:validation:test cohorts (ratio of 5:1:4). An expert-radiologist manually delineated all rectal tumours to serve as training input. Using the T2-weighted, diffusion-weighted imaging (DWI)-b1000 and apparent diffusion coefficient (ADC)-maps from the training+validation cohorts, an attention-gated u-net was trained, with optimal hyperparameters (including learning rate, optimizer algorithm and batch-size) determined via grid search. A second expert-radiologist independently re-segmented all patients in the test cohort to calculate inter-reader agreement using the dice similarity coefficient (DSC). Agreement (DSC) between the trained network and the expert-segmentations was compared to the expert inter-reader agreement (serving as the standard of reference).
Results
Average DSC between expert-readers was 0.75 (±0.18). Average DSC for the network-generated segmentations was 0.64 (±0.22) and 0.59 (±0.22) compared to expert-reader 1 and 2, respectively. When excluding poor quality (score ≤ 2) scans, DSC between the network and expert-readers increased to 0.68 (±0.17) and 0.63 (±0.17).
Conclusion
Despite large variations in scan protocol/image quality, deep learning networks achieved a promising overall performance to automatically segment rectal tumours on MRI. Better results were achieved after exclusion of poor-quality scans, with agreement levels up to 0.63-0.68 (versus 0.75 between expert-radiologists).
Video-on-demand
Author of 1 Presentation
SS 5.4 - Deep learning for fully automated segmentation of rectal tumours on MRI in a multicenter setting (ID 849)
Abstract
Purpose
To explore the performance of deep learning to perform fully automated segmentation of rectal tumors on baseline MRI using a multicenter dataset (with variations in acquisition protocols and scan quality reflecting daily practice).
Material and methods
Baseline staging MRIs of 355 patients (from 6 centers) were analyzed. Overall scan quality was assessed using a 5-point Likert score (0=poor) to 4=excellent). Data were randomly split into train:validation:test cohorts (ratio of 5:1:4). An expert-radiologist manually delineated all rectal tumours to serve as training input. Using the T2-weighted, diffusion-weighted imaging (DWI)-b1000 and apparent diffusion coefficient (ADC)-maps from the training+validation cohorts, an attention-gated u-net was trained, with optimal hyperparameters (including learning rate, optimizer algorithm and batch-size) determined via grid search. A second expert-radiologist independently re-segmented all patients in the test cohort to calculate inter-reader agreement using the dice similarity coefficient (DSC). Agreement (DSC) between the trained network and the expert-segmentations was compared to the expert inter-reader agreement (serving as the standard of reference).
Results
Average DSC between expert-readers was 0.75 (±0.18). Average DSC for the network-generated segmentations was 0.64 (±0.22) and 0.59 (±0.22) compared to expert-reader 1 and 2, respectively. When excluding poor quality (score ≤ 2) scans, DSC between the network and expert-readers increased to 0.68 (±0.17) and 0.63 (±0.17).
Conclusion
Despite large variations in scan protocol/image quality, deep learning networks achieved a promising overall performance to automatically segment rectal tumours on MRI. Better results were achieved after exclusion of poor-quality scans, with agreement levels up to 0.63-0.68 (versus 0.75 between expert-radiologists).