Table 1 Performance comparison of CGS-Net models vs. baseline models for cancer segmentation. The CGS-Net models outperformed their single-input baseline models. These results were seen across AUC and cancer Dice scores for all model sizes and architectures.
From: Context-guided segmentation for histopathologic cancer segmentation
Model | Parameters | AUC | CGS-Net Improvement | Cancer Dice score | CGS-Net Improvement |
|---|---|---|---|---|---|
Single MiT-B1 | 15.14 M | \(0.976 \pm 0.0008\) | - | \(0.6677 \pm 0.108\) | - |
CGS-Net MiT-B1 | 30.08 M | \(\mathbf {0.9806} \pm \mathbf {0.0002}\) | 0.47% | \(\mathbf {0.7004} \pm \mathbf {0.1228}\) | 4.09% |
Single MiT-B2 | 26.18 M | \(0.972 \pm 0.002\) | - | \(0.5719 \pm 0.0326\) | - |
CGS-Net MiT-B2 | 52.17 M | \(\mathbf {0.979} \pm \mathbf {0.0006}\) | 0.72% | \(\mathbf {0.6036} \pm \mathbf {0.0098}\) | 5.54% |
Single SwinV2 Tiny | 29.56 M | \(0.971 \pm 0.0009\) | - | \(0.5735 \pm 0.04\) | - |
CGS-Net SwinV2 Tiny | 60.81 M | \(\mathbf {0.974} \pm \mathbf {0.0015}\) | 0.31% | \(\mathbf {0.6126} \pm \mathbf {0.0035}\) | 6.81% |
Single SwinV2 Small | 50.92 M | \(0.967 \pm 0.0097\) | - | \(0.5631 \pm 0.029\) | - |
CGS-Net SwinV2 Small | 103.54 M | \(\mathbf {0.976} \pm \mathbf {0.0008}\) | 0.93% | \(\mathbf {0.5936} \pm \mathbf {0.0066}\) | 5.42% |
