Skip to main content

Table 1 Model performance/reliability measured by SDC/HD with/without curation, per cost function and for set window versus patient-specific window

From: Strategies to improve deep learning-based salivary gland segmentation

   Set sizes (SMG/PG) SDC HD
   Train Validation Test SMG PG SMG PG
Data quality
Without augmentation
 Train data Test data        
 Clinical Clinical 18/18 2/2 4/4 .68 ± .06 .68 ± .05 17.6 ± 1.5 24.7 ± 5.1
 Curated Curated 18/18 2/2 4/4 .66 ± .07 .68 ± .04 23.4 ± 1.3 28.1 ± 4.3
With augmentation
 Train data Test data        
 Clinical Clinical 90/90 2/2 4/4 .67 ± .06 .69 ± .06 13.6 ± 1.3 24.8 ± 5.3
 Curated Curated 90/90 2/2 4/4 .67 ± .07 .69 ± .04 12.0 ± 1.5 21.8 ± 4.6
Cost functions
 SDC   90/90 10/10 20/20 .71 ± .06 .71 ± .06 6.9 ± 1.5 17.3 ± 6.4
 SDC(0.5)   90/90 10/10 20/20 .71 ± .06 .71 ± .06 9.0 ± 3.0 17.4 ± 6.8
 SDC(0.05)   90/90 10/10 20/20 .70 ± .06 .71 ± .06 6.6 ± 1.6 16.6 ± 7.0
 SDC + HD   90/90 10/10 20/20 .70 ± .05   7.6 ± 2.4  
Patient-specific windowing
 Set window   940/1024 94/114 188/227 .86 ± .07 .85 ± .05 4.5 ± 1.9 8.1 ± 3.8
 Patient-specific window   940/1024 94/114 188/227 .87 ± .05 .87 ± .04 4.1 ± 1.6 7.7 ± 3.8