Complete medical imaging workspace dump (weights, deliverables, papers, notebooks, code, figures)
eded815 verified |
Download test_set_evaluation_report.md from unknownBrain/BioHCL-SSSM-Complete-Workspace: direct link, hf CLI and curl.
- Browser
- Download file 2.38 kB
-
https://huggingface.co/unknownBrain/BioHCL-SSSM-Complete-Workspace/resolve/main/test_set_evaluation_report.md
- Command line
-
hf download hf://unknownBrain/BioHCL-SSSM-Complete-Workspace/test_set_evaluation_report.md
-
curl -L -o test_set_evaluation_report.md https://huggingface.co/unknownBrain/BioHCL-SSSM-Complete-Workspace/resolve/main/test_set_evaluation_report.md
2.38 kB
| # Held-Out Sequestered Test Set Evaluation Report | |
| This report presents the diagnostic evaluation on the official held-out test split of the Chest X-Ray Pneumonia dataset ($N = 624$ images: $234$ NORMAL, $390$ PNEUMONIA). These images were strictly sequestered and never seen during cross-validation or hyperparameter selection. | |
| --- | |
| ## 1. Held-Out Test Set Performance Matrix | |
| | Model Architecture | Test Accuracy (%) | Test Precision | Test Recall (Sensitivity) | Test F1-Score | Test ROC-AUC | Parameters (M) | Parameter Reduction vs Baseline | | |
| | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | |
| | **ResNet-50 Baseline** | 84.46% | 0.8033 | 0.9949 | 0.8889 | 0.9605 | 23.51 M | Baseline (0%) | | |
| | **BioHCL-SSSM (No BioPrior Ablation)** | 82.37% | 0.7811 | **0.9974** | 0.8761 | 0.9506 | 14.01 M | -40.4% | | |
| | **Student BioHCL-SSSM (Full Model)** | **87.50%** | **0.8362** | 0.9949 | **0.9087** | 0.9391 | **14.01 M** | **-40.4% (-9.50M params)** | | |
| ### Key Diagnostic Highlights | |
| * **Accuracy Improvement:** BioHCL-SSSM surpasses ResNet-50 by **+3.04 percentage points** ($87.50\%$ vs $84.46\%$). | |
| * **F1 Improvement:** Harmonic mean rises from **0.8889 to 0.9087** ($+0.0198$). | |
| * **Model Footprint:** BioHCL-SSSM reduces parameter count by **40.4%** ($14.01\text{M}$ vs $23.51\text{M}$). | |
| * **Bio-Prior Efficacy:** Ablating the biological prior vector drops test accuracy by **5.13 percentage points** ($87.50\% \rightarrow 82.37\%$) and F1 by **-0.0326**, verifying that physics-informed biological inductive biases improve test generalization. | |
| --- | |
| ## 2. Test Set Visualizations & Confusion Matrices | |
| ````carousel | |
| ### 1. Student BioHCL-SSSM Test Confusion Matrix | |
|  | |
| <!-- slide --> | |
| ### 2. Student BioHCL-SSSM Test ROC Curve | |
|  | |
| <!-- slide --> | |
| ### 3. Baseline ResNet-50 Test Confusion Matrix | |
|  | |
| <!-- slide --> | |
| ### 4. Baseline ResNet-50 Test ROC Curve | |
|  | |
| <!-- slide --> | |
| ### 5. ROC & Precision-Recall Comparison (Ablation) | |
|  | |
| <!-- slide --> | |
| ### 6. BioPrior Ablated Model Test Confusion Matrix | |
|  | |
| ```` | |