Table 3:
Detecting cancer: Lesion-level evaluation on cohorts C1-test and cohort C2.
| Evaluation on cohort C1-test | ||||||||||
| Patients with cancer lesion volumes ≥ 250 mm3 N = 37, number of cancerous lesions = 44 | ||||||||||
|
| ||||||||||
| Model | ROC AUC | PR AUC | Sensitivity | Specificity | Precision | NPV | F1 Score | Dice | Accuracy | Final Rank |
|
| ||||||||||
| SPCNet | 0.82±0.30 | 0.76±0.36 | 0.62±0.47 | 0.75±0.30 | 0.44±0.41 | 0.84±0.22 | 0.48±0.40 | 0.27±0.26 | 0.72±0.24 | 5 |
| U-Net | 0.82±0.34 | 0.81±0.34 | 0.84±0.35 | 0.30±0.31 | 0.35±0.27 | 0.57±0.47 | 0.45±0.25 | 0.30±0.20 | 0.43±0.26 | 8 |
| BrU-Net | 0.80± 0.32 | 0.78±0.34 | 0.82±0.35 | 0.39±0.38 | 0.37±0.29 | 0.57±0.46 | 0.47±0.28 | 0.27±0.17 | 0.52±0.27 | 9 |
|
| ||||||||||
| CorrSigNIA | 0.81±0.31 | 0.73±0.38 | 0.73±0.43 | 0.72±0.33 | 0.50±0.41 | 0.86±0.25 | 0.55±0.39 | 0.30±0.24 | 0.72±0.26 | 1 |
| SPCNet + CR-3-S | 0.79±0.31 | 0.72±0.38 | 0.62±0.46 | 0.80±0.21 | 0.45±0.39 | 0.88±0.15 | 0.49±0.38 | 0.27±0.23 | 0.75±0.17 | 4 |
| SPCNet + CR-123-I | 0.79±0.33 | 0.72±0.37 | 0.66±0.47 | 0.77±0.29 | 0.48±0.42 | 0.87±0.21 | 0.52±0.41 | 0.29±0.24 | 0.75±0.23 | 2 |
| SPCNet + CR-123-S | 0.73±0.35 | 0.70±0.38 | 0.55±0.48 | 0.82±0.24 | 0.41±0.41 | 0.83±0.21 | 0.44±0.40 | 0.27±0.24 | 0.75±0.20 | 11 |
| SPCNet + CR-4-I | 0.80±0.34 | 0.77±0.36 | 0.58±0.47 | 0.80±0.30 | 0.44±0.42 | 0.82±0.25 | 0.47±0.42 | 0.28±0.23 | 0.73±0.24 | 7 |
| SPCNet + CR-4-S | 0.79±0.34 | 0.75±0.37 | 0.69±0.44 | 0.72±0.32 | 0.50±0.41 | 0.82±0.28 | 0.54±0.39 | 0.30±0.23 | 0.71±0.27 | 3 |
| U-Net + CR-3-I | 0.80±0.34 | 0.77±0.34 | 0.89±0.29 | 0.27±0.33 | 0.34±0.23 | 0.46±0.48 | 0.47±0.24 | 0.29±0.17 | 0.44±0.26 | 10 |
| U-Net + CR-123-I | 0.80±0.31 | 0.72±0.38 | 0.88±0.32 | 0.41±0.37 | 0.40±0.29 | 0.65±0.46 | 0.51±0.28 | 0.31±0.21 | 0.55±0.28 | 6 |
| BrU-Net + CR-3-I | 0.77±0.38 | 0.75±0.38 | 0.88±0.32 | 0.10±0.25 | 0.30±0.22 | 0.20±0..9 | 0.42±0.23 | 0.27±0.17 | 0.32±0.23 | 13 |
| BrU-Net + CR-123-I | 0.72±0.39 | 0.66±0.40 | 0.69±0.44 | 0.58±0.38 | 0.43±0.38 | 0.70±0.38 | 0.49±0.37 | 0.27±0.23 | 0.62±0.30 | 12 |
|
| ||||||||||
| Evaluation on cohort C2 | ||||||||||
| Patients with cancer N = 147, number of cancerous lesions = 189 | ||||||||||
|
| ||||||||||
| Model | ROC AUC | PR AUC | Sensitivity | Specificity | Precision | NPV | F1 Score | Dice | Accuracy | Final Rank |
|
| ||||||||||
| SPCNet | 0.75±0.36 | 0.76±0.35 | 0.50±0.47 | 0.77±0.36 | 0.42±0.43 | 0.69±0.33 | 0.43±0.42 | 0.25±0.25 | 0.69±0.24 | 5 |
| U-Net | 0.78±0.33 | 0.77±0.34 | 0.76±0.39 | 0.57±0.39 | 0.50±0.35 | 0.66±0.41 | 0.56±0.33 | 0.33±0.23 | 0.63±0.26 | 3 |
| BrU-Net | 0.79± 0.35 | 0.79±0.34 | 0.89±0.29 | 0.27±0.36 | 0.41±0.26 | 0.41±0.47 | 0.53±0.25 | 0.33±0.20 | 0.48±0.26 | 4 |
|
| ||||||||||
| CorrSigNIA | 0.81±0.31 | 0.79±0.33 | 0.58±0.46 | 0.76±0.36 | 0.49±0.43 | 0.71±0.33 | 0.49±0.41 | 0.30±0.26 | 0.71±0.25 | 1 |
| SPCNet + CR-123-I | 0.77±0.33 | 0.76±0.34 | 0.55±0.47 | 0.81±0.33 | 0.47±0.44 | 0.75±0.29 | 0.48±0.42 | 0.30±0.27 | 0.74±0.22 | 2 |
| U-Net + CR-3-I | 0.49±0.43 | 0.54±0.39 | 0.40±0.46 | 0.59±0.40 | 0.30±0.38 | 0.56±0.36 | 0.31±0.37 | 0.19±0.22 | 0.53±0.29 | 9 |
| U-Net + CR-123-I | 0.67±0.38 | 0.66±0.38 | 0.52±0.47 | 0.67±0.39 | 0.41±0.41 | 0.65±0.35 | 0.42±0.39 | 0.24±0.24 | 0.62±0.28 | 7 |
| BrU-Net + CR-3-I | 0.81±0.33 | 0.81±0.33 | 0.90±0.27 | 0.12±0.26 | 0.36±0.20 | 0.22±0.40 | 0.49±0.20 | 0.32±0.19 | 0.39±0.21 | 6 |
| BrU-Net + CR-123-I | 0.61±0.40 | 0.62±0.39 | 0.41±0.46 | 0.78±0.34 | 0.37±0.44 | 0.68±0.30 | 0.36±0.41 | 0.21±0.25 | 0.66±0.25 | 7 |
Mean and standard deviation values for each metric in each cohort is reported. Column “Final rank” represents rank of model based on sum of individual metric ranks (detailed table with individual metric ranks in Supplementary table S1). The first three rows above the dotted lines in both cohorts represent the MRI-only models, whereas the rows below the dotted lines represent the MRI + correlated features based models. CorrSigNIA outperforms all models in cancer detection.