Initial release: compressed CheXNet + surgical correction artifacts

.gitattributes

@@ -33,3 +33,10 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+figures/fig2_sparse_layers.png filter=lfs diff=lfs merge=lfs -text
+figures/fig3_mutual_exclusivity.pdf filter=lfs diff=lfs merge=lfs -text
+figures/fig3_mutual_exclusivity.png filter=lfs diff=lfs merge=lfs -text
+figures/fig5_surgery.png filter=lfs diff=lfs merge=lfs -text
+figures/fig6_treatment.png filter=lfs diff=lfs merge=lfs -text
+figures/fig8_clinical_report.pdf filter=lfs diff=lfs merge=lfs -text
+figures/fig8_clinical_report.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,215 @@
+---
+license: mit
+tags:
+- medical-imaging
+- chest-xray
+- chexnet
+- densenet121
+- model-compression
+- mechanistic-interpretability
+- xai
+datasets:
+- nih-chest-xray
+library_name: pytorch
+pipeline_tag: image-classification
+---
+
+# Lossless Mechanistic Compression and Surgical Correction of Medical Imaging Models
+
+Artifacts for the paper by Yeonseong Cynn (River Lab, May 2026).
+
+## Summary
+
+A compressed CheXNet (DenseNet121) at **51.43% parameter reduction**
+(6,966,034 → 3,383,248) with mean AUROC preserved within sampling noise
+on n=1045 NIH ChestX-ray14 test images (Δ +0.0004, per-pathology max
+|Δ| = 0.0033). Output identity to numerical precision
+(max |Δ logit| < 5×10⁻⁶).
+
+The compressed model exposes classifier channels at a granularity that
+makes mechanistic interventions practical:
+
+- **Surgical correction**: 5-channel classifier weight zeroing softly
+  reduces a target false-positive probability with bounded side effects.
+- **Mutual exclusivity insight**: 89 of 100 polarized classifier channels
+  are not architectural conflicts but bipolar discriminative axes
+  exploiting label mutual exclusivity (Jaccard < 0.1).
+- **Cost-aware operations**: threshold calibration and minimal retraining
+  routed by a decision system per pathology.
+- **Clinical report auto-generation**: combining channel-level evidence,
+  Grad-CAM region mapping, and mutual-exclusivity exclusion.
+
+## Files
+
+### Weights
+| File | Size | Description |
+|---|---|---|
+| `compressed_model.pt` | 14.2 MB | Compressed CheXNet backbone + classifier (3.38M params) |
+| `classifier_finetuned.pt` | 75 KB | Optional fine-tuned classifier head (18K params) |
+
+### Code
+| File | Description |
+|---|---|
+| `inference.py` | Minimal CLI inference (load + forward) |
+| `requirements.txt` | Pip dependencies |
+
+### Metrics (JSON)
+- `metrics/baseline_vs_compressed.json` — Per-pathology AUROC (baseline vs compressed, n=1045)
+- `metrics/eval_nih_weights.json` — All 5 torchxrayvision DenseNet121 checkpoints on NIH test
+- `metrics/analyze_binary_axis.json` — Pathology independence + Jaccard matrix
+- `metrics/q_conflict_legitimacy.json` — Polarized channel legitimacy classification
+- `metrics/surgery_channel_ablation.json` — Surgical correction K-sweep
+- `metrics/apply_threshold_calibration.json` — Per-class Youden threshold + F1/Recall
+- `metrics/minimal_retrain.json`, `minimal_retrain_v2.json` — Classifier-head fine-tune costs
+
+### Figures (paper)
+- `figures/fig1_compression.{pdf,png}` — Headline numbers
+- `figures/fig2_sparse_layers.{pdf,png}` — Per-block sparsity
+- `figures/fig3_mutual_exclusivity.{pdf,png}` — Jaccard ↔ channel-usage mirror
+- `figures/fig4_legitimacy.{pdf,png}` — Polarized-channel legitimacy
+- `figures/fig5_surgery.{pdf,png}` — Surgical-correction K-sweep
+- `figures/fig6_treatment.{pdf,png}` — Per-pathology treatment recommendations
+- `figures/fig7_minimal_retrain.{pdf,png}` — Fine-tuning + threshold calibration
+- `figures/fig8_clinical_report.{pdf,png}` — Clinical report sample
+
+## Setup
+
+```bash
+pip install -r requirements.txt
+```
+
+## Usage
+
+### Loading and inference
+
+```bash
+python inference.py path/to/xray.png
+python inference.py path/to/xray.png --classifier classifier_finetuned.pt
+```
+
+### Loading the model in your own code
+
+```python
+import torch
+import torch.nn as nn
+import torchxrayvision as xrv
+
+model = xrv.models.DenseNet(weights="densenet121-res224-all").eval()
+ckpt = torch.load("compressed_model.pt", weights_only=False)
+for block_idx in [1, 2, 3, 4]:
+    block = getattr(model.features, f"denseblock{block_idx}")
+    block_alive = ckpt["alive_per_block"][block_idx]
+    for dl_key, n_alive in block_alive.items():
+        i = int(dl_key[2:])
+        L = getattr(block, f"denselayer{i}")
+        in_ch = L.conv1.in_channels
+        L.conv1 = nn.Conv2d(in_ch, n_alive, 1, bias=True).eval()
+        L.norm2 = nn.BatchNorm2d(n_alive, eps=L.norm2.eps).eval()
+        L.conv2 = nn.Conv2d(n_alive, 32, 3, padding=1, bias=False).eval()
+model.load_state_dict(ckpt["state_dict"])
+for block_idx in [1, 2, 3, 4]:
+    block = getattr(model.features, f"denseblock{block_idx}")
+    for i in range(1, {1:6, 2:12, 3:24, 4:16}[block_idx] + 1):
+        getattr(block, f"denselayer{i}").norm2 = nn.Identity()
+
+# Optional fine-tuned classifier
+cls_ft = nn.Linear(1024, 18)
+cls_ft.load_state_dict(torch.load("classifier_finetuned.pt", weights_only=True))
+model.classifier = cls_ft
+model.eval()
+```
+
+## Verification
+
+### NIH ChestX-ray14 official test split (1045 images)
+
+| Configuration | Parameters | Mean AUROC | Latency (ms/image) |
+|---|---|---|---|
+| Baseline (`densenet121-res224-all`) | 6,966,034 | 0.7781 | 15.17 |
+| **Compressed** | **3,383,248 (-51.43%)** | **0.7785 (+0.0004)** | **14.73 (-2.9%)** |
+
+Per-pathology max |Δ AUROC| = 0.0033 (Emphysema +); all within sampling noise.
+
+### Choice of baseline checkpoint
+
+We compared all 5 `torchxrayvision` DenseNet121 checkpoints on the same
+NIH test subset. The multi-source `all` is the strongest:
+
+| Checkpoint | Mean AUROC |
+|---|---|
+| **`densenet121-res224-all`** | **0.7781** |
+| `densenet121-res224-nih` | 0.7524 |
+| `densenet121-res224-chex` | 0.7425 |
+| `densenet121-res224-mimic_ch` | 0.7178 |
+| `densenet121-res224-mimic_nb` | 0.7049 |
+
+Higher published NIH-only DenseNet121 numbers (e.g., 0.84) come from
+corpus-specific hyperparameter and augmentation tuning not part of the
+open `torchxrayvision` release.
+
+### Threshold calibration (Youden-J)
+
+The default decision threshold 0.5 is overly conservative for this
+multi-label model. Per-class Youden-J on a held-out validation set
+shifts the cohort-average operating point:
+
+| Setting | Mean F1 | Mean Recall |
+|---|---|---|
+| Default threshold 0.5 | 0.127 | 0.111 |
+| Youden-J calibrated | **0.20** | **0.78** |
+
+**Caveat**: this trades precision for recall sharply. Best-performing
+classes (Cardiomegaly: precision 1.0 → 0.11, F1 0.57 → 0.20; Mass: F1
+0.25 → 0.07) are degraded. F1 average is dominated by previously
+zero-recall classes (Infiltration, Atelectasis). For deployment,
+F1-optimal thresholds or explicit clinical precision floors are
+preferable.
+
+### Surgical correction (representative Cardiomegaly false positive)
+
+| K (channels zeroed) | Target prob | TP loss | Other 13 pathology AUROC Δ |
+|---|---|---|---|
+| 0 (baseline) | 0.89 | — | — |
+| 5 | 0.76 | 0 | exactly 0 |
+| 10 | 0.67 | 4 | exactly 0 |
+| 20 | 0.54 | 7 | exactly 0 |
+
+At K=5 the decision (threshold 0.5) is **not flipped**; the correction
+is a soft probability reduction, not a hard decision change. K=20
+crosses the boundary but loses 7 true positives. Treat surgical
+correction as a confidence-shaping tool, not a binary error eraser.
+
+The exact-zero AUROC isolation guarantee on the other 13 pathologies
+holds by construction (only one classifier row is modified).
+
+## Method Disclosure
+
+Compression method specifics are proprietary; the foundational procedure
+is covered by Korean patent applications. The released artifacts (weights,
+inference code, downstream analysis scripts) are sufficient for
+reproduction of the reported results.
+
+## Base Model
+
+[torchxrayvision densenet121-res224-all](https://github.com/mlmed/torchxrayvision)
+(DenseNet121 trained on NIH ChestX-ray14, CheXpert, MIMIC-CXR, PadChest).
+
+## Citation
+
+```bibtex
+@misc{cynn2026chexnet,
+  title={Lossless Mechanistic Compression and Surgical Correction of Medical Imaging Models},
+  author={Cynn, Yeonseong},
+  year={2026},
+  note={Manuscript in preparation}
+}
+```
+
+## License
+
+MIT for the released code; the underlying compression method is
+proprietary (see Method Disclosure above).
+
+## Contact
+
+For questions or commercial inquiries: whitepep@gmail.com

classifier_finetuned.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb003a949793d1e60f7c939d45a2c6b14b4b1868a8838687351a6872ffaa6e8d
+size 75430

compressed_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:338f9366a1db6cff074a2e2e882747a6d9a5e8b09fcf7be43880cebf38db2de2
+size 14165486

figures/fig3_mutual_exclusivity.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de9122621761c12e077999e0fcde4b98f9f38c6923b5410739f3c1d8f2a9f310
+size 187145

figures/fig8_clinical_report.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a9e9e602b6756e1df45d513fd96d75910725dac7b09cbdb8d4cfedf4a061e18
+size 314896

inference.py

@@ -0,0 +1,110 @@
+"""
+CheXNet Compressed Model — Inference
+
+Loads the compressed model (optionally with the fine-tuned classifier head)
+and predicts pathology probabilities for a chest X-ray image.
+
+Usage:
+    python inference.py <path/to/image.png>
+    python inference.py <path/to/image.png> --classifier classifier_finetuned.pt
+"""
+import sys
+sys.dont_write_bytecode = True  # avoid creating __pycache__ in the release dir
+
+import argparse
+import os
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchxrayvision as xrv
+from PIL import Image
+
+
+def build_compressed_model(ckpt_path, device, classifier_ft_path=None):
+    """Load baseline CheXNet then reshape to compressed dimensions and load weights."""
+    model = xrv.models.DenseNet(weights="densenet121-res224-all").to(device).eval()
+    ckpt = torch.load(ckpt_path, map_location=device, weights_only=False)
+    alive = ckpt["alive_per_block"]
+    for block_idx in [1, 2, 3, 4]:
+        block = getattr(model.features, f"denseblock{block_idx}")
+        block_alive = alive.get(block_idx, alive.get(f"block{block_idx}", {}))
+        for dl_key, n_alive in block_alive.items():
+            i = int(dl_key[2:]) if dl_key.startswith("dl") else int(dl_key)
+            L = getattr(block, f"denselayer{i}")
+            in_ch = L.conv1.in_channels
+            L.conv1 = nn.Conv2d(in_ch, n_alive, 1, bias=True).to(device).eval()
+            L.norm2 = nn.BatchNorm2d(n_alive, eps=L.norm2.eps).to(device).eval()
+            L.conv2 = nn.Conv2d(n_alive, 32, 3, padding=1, bias=False).to(device).eval()
+    model.load_state_dict(ckpt["state_dict"])
+    # intermediate batch norms are folded; replace with identity
+    for block_idx in [1, 2, 3, 4]:
+        block = getattr(model.features, f"denseblock{block_idx}")
+        n_layers = {1: 6, 2: 12, 3: 24, 4: 16}[block_idx]
+        for i in range(1, n_layers + 1):
+            getattr(block, f"denselayer{i}").norm2 = nn.Identity()
+
+    if classifier_ft_path and os.path.exists(classifier_ft_path):
+        cls_ft = nn.Linear(1024, 18).to(device)
+        cls_ft.load_state_dict(torch.load(classifier_ft_path,
+                                          map_location=device, weights_only=True))
+        model.classifier = cls_ft
+
+    model.eval()
+    return model
+
+
+def preprocess(img_np):
+    """xrv normalization: scale to [-1024, 1024] and resize to 224x224."""
+    arr = img_np.astype(np.float32)
+    arr = (arr - arr.min()) / max(arr.max() - arr.min(), 1) * 2048 - 1024
+    pil = Image.fromarray(arr, mode="F").resize((224, 224), Image.BILINEAR)
+    return np.array(pil)
+
+
+def predict(model, image_path, device, top_n=5):
+    img = Image.open(image_path)
+    if img.mode != "L":
+        img = img.convert("L")
+    img_proc = preprocess(np.array(img))
+    t = torch.tensor(img_proc, dtype=torch.float32).unsqueeze(0).unsqueeze(0).to(device)
+    with torch.no_grad():
+        feat = model.features(t)
+        feat = F.relu(feat, inplace=False)
+        pooled = F.adaptive_avg_pool2d(feat, (1, 1)).flatten(1)
+        logits = model.classifier(pooled)[0]
+        probs = torch.sigmoid(logits).cpu().numpy()
+    pathologies = list(model.pathologies)
+    return sorted(zip(pathologies, probs.tolist()), key=lambda x: -x[1])[:top_n]
+
+
+def main():
+    parser = argparse.ArgumentParser(description="CheXNet compressed model inference")
+    parser.add_argument("image_path", help="Path to chest X-ray image (PNG)")
+    parser.add_argument("--ckpt", default="compressed_model.pt",
+                        help="Compressed model checkpoint")
+    parser.add_argument("--classifier", default="classifier_finetuned.pt",
+                        help="Optional fine-tuned classifier head")
+    parser.add_argument("--top", type=int, default=5, help="Number of predictions to show")
+    args = parser.parse_args()
+
+    here = os.path.dirname(os.path.abspath(__file__))
+    ckpt_path = args.ckpt if os.path.isabs(args.ckpt) else os.path.join(here, args.ckpt)
+    cls_path = args.classifier if os.path.isabs(args.classifier) else os.path.join(here, args.classifier)
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    print(f"Device: {device}")
+
+    model = build_compressed_model(ckpt_path, device, cls_path)
+    n_params = sum(p.numel() for p in model.parameters())
+    print(f"Model parameters: {n_params:,}")
+
+    ranked = predict(model, args.image_path, device, top_n=args.top)
+    print(f"\nTop-{args.top} predictions for {args.image_path}:")
+    for pathology, prob in ranked:
+        print(f"  {pathology:<28}  {prob:.4f}")
+
+
+if __name__ == "__main__":
+    main()

metrics/analyze_binary_axis.json

@@ -0,0 +1,324 @@
+{
+  "pathology_independence_score": {
+    "Atelectasis": 0.9553384421051226,
+    "Consolidation": 0.9511092287725019,
+    "Infiltration": 0.9473420384158613,
+    "Pneumothorax": 0.9851201592772656,
+    "Edema": 0.9658843078521789,
+    "Emphysema": 0.9924951267056531,
+    "Fibrosis": 0.9632083381606142,
+    "Effusion": 0.9271281593875844,
+    "Pneumonia": 0.984717957807288,
+    "Pleural_Thickening": 0.9555283714532556,
+    "Cardiomegaly": 0.9483568960925801,
+    "Nodule": 0.9583928226418525,
+    "Mass": 0.9565895170805814,
+    "Hernia": 0.9917770617018739
+  },
+  "jaccard_matrix": {
+    "Atelectasis__Consolidation": 0.0963855421686747,
+    "Atelectasis__Infiltration": 0.13194444444444445,
+    "Atelectasis__Pneumothorax": 0.012658227848101266,
+    "Atelectasis__Edema": 0.0,
+    "Atelectasis__Emphysema": 0.0,
+    "Atelectasis__Fibrosis": 0.010869565217391304,
+    "Atelectasis__Effusion": 0.17424242424242425,
+    "Atelectasis__Pneumonia": 0.02857142857142857,
+    "Atelectasis__Pleural_Thickening": 0.04819277108433735,
+    "Atelectasis__Cardiomegaly": 0.03773584905660377,
+    "Atelectasis__Nodule": 0.0,
+    "Atelectasis__Mass": 0.04,
+    "Atelectasis__Hernia": 0.0,
+    "Consolidation__Atelectasis": 0.0963855421686747,
+    "Consolidation__Infiltration": 0.05172413793103448,
+    "Consolidation__Pneumothorax": 0.02631578947368421,
+    "Consolidation__Edema": 0.06451612903225806,
+    "Consolidation__Emphysema": 0.0,
+    "Consolidation__Fibrosis": 0.0196078431372549,
+    "Consolidation__Effusion": 0.09615384615384616,
+    "Consolidation__Pneumonia": 0.0,
+    "Consolidation__Pleural_Thickening": 0.06976744186046512,
+    "Consolidation__Cardiomegaly": 0.029850746268656716,
+    "Consolidation__Nodule": 0.09302325581395349,
+    "Consolidation__Mass": 0.08823529411764706,
+    "Consolidation__Hernia": 0.0,
+    "Infiltration__Atelectasis": 0.13194444444444445,
+    "Infiltration__Consolidation": 0.05172413793103448,
+    "Infiltration__Pneumothorax": 0.01834862385321101,
+    "Infiltration__Edema": 0.039603960396039604,
+    "Infiltration__Emphysema": 0.0,
+    "Infiltration__Fibrosis": 0.06896551724137931,
+    "Infiltration__Effusion": 0.1625,
+    "Infiltration__Pneumonia": 0.03,
+    "Infiltration__Pleural_Thickening": 0.02608695652173913,
+    "Infiltration__Cardiomegaly": 0.09302325581395349,
+    "Infiltration__Nodule": 0.05309734513274336,
+    "Infiltration__Mass": 0.009259259259259259,
+    "Infiltration__Hernia": 0.0,
+    "Pneumothorax__Atelectasis": 0.012658227848101266,
+    "Pneumothorax__Consolidation": 0.02631578947368421,
+    "Pneumothorax__Infiltration": 0.01834862385321101,
+    "Pneumothorax__Edema": 0.0,
+    "Pneumothorax__Emphysema": 0.05,
+    "Pneumothorax__Fibrosis": 0.0,
+    "Pneumothorax__Effusion": 0.0,
+    "Pneumothorax__Pneumonia": 0.0,
+    "Pneumothorax__Pleural_Thickening": 0.0,
+    "Pneumothorax__Cardiomegaly": 0.017543859649122806,
+    "Pneumothorax__Nodule": 0.02857142857142857,
+    "Pneumothorax__Mass": 0.04,
+    "Pneumothorax__Hernia": 0.0,
+    "Edema__Atelectasis": 0.0,
+    "Edema__Consolidation": 0.06451612903225806,
+    "Edema__Infiltration": 0.039603960396039604,
+    "Edema__Pneumothorax": 0.0,
+    "Edema__Emphysema": 0.0,
+    "Edema__Fibrosis": 0.029411764705882353,
+    "Edema__Effusion": 0.043010752688172046,
+    "Edema__Pneumonia": 0.07692307692307693,
+    "Edema__Pleural_Thickening": 0.0,
+    "Edema__Cardiomegaly": 0.15555555555555556,
+    "Edema__Nodule": 0.034482758620689655,
+    "Edema__Mass": 0.0,
+    "Edema__Hernia": 0.0,
+    "Emphysema__Atelectasis": 0.0,
+    "Emphysema__Consolidation": 0.0,
+    "Emphysema__Infiltration": 0.0,
+    "Emphysema__Pneumothorax": 0.05,
+    "Emphysema__Edema": 0.0,
+    "Emphysema__Fibrosis": 0.0,
+    "Emphysema__Effusion": 0.010526315789473684,
+    "Emphysema__Pneumonia": 0.0,
+    "Emphysema__Pleural_Thickening": 0.037037037037037035,
+    "Emphysema__Cardiomegaly": 0.0,
+    "Emphysema__Nodule": 0.0,
+    "Emphysema__Mass": 0.0,
+    "Emphysema__Hernia": 0.0,
+    "Fibrosis__Atelectasis": 0.010869565217391304,
+    "Fibrosis__Consolidation": 0.0196078431372549,
+    "Fibrosis__Infiltration": 0.06896551724137931,
+    "Fibrosis__Pneumothorax": 0.0,
+    "Fibrosis__Edema": 0.029411764705882353,
+    "Fibrosis__Emphysema": 0.0,
+    "Fibrosis__Effusion": 0.06422018348623854,
+    "Fibrosis__Pneumonia": 0.0,
+    "Fibrosis__Pleural_Thickening": 0.11627906976744186,
+    "Fibrosis__Cardiomegaly": 0.028985507246376812,
+    "Fibrosis__Nodule": 0.11363636363636363,
+    "Fibrosis__Mass": 0.02631578947368421,
+    "Fibrosis__Hernia": 0.0,
+    "Effusion__Atelectasis": 0.17424242424242425,
+    "Effusion__Consolidation": 0.09615384615384616,
+    "Effusion__Infiltration": 0.1625,
+    "Effusion__Pneumothorax": 0.0,
+    "Effusion__Edema": 0.043010752688172046,
+    "Effusion__Emphysema": 0.010526315789473684,
+    "Effusion__Fibrosis": 0.06422018348623854,
+    "Effusion__Pneumonia": 0.021505376344086023,
+    "Effusion__Pleural_Thickening": 0.0784313725490196,
+    "Effusion__Cardiomegaly": 0.17699115044247787,
+    "Effusion__Nodule": 0.05714285714285714,
+    "Effusion__Mass": 0.052083333333333336,
+    "Effusion__Hernia": 0.010526315789473684,
+    "Pneumonia__Atelectasis": 0.02857142857142857,
+    "Pneumonia__Consolidation": 0.0,
+    "Pneumonia__Infiltration": 0.03,
+    "Pneumonia__Pneumothorax": 0.0,
+    "Pneumonia__Edema": 0.07692307692307693,
+    "Pneumonia__Emphysema": 0.0,
+    "Pneumonia__Fibrosis": 0.0,
+    "Pneumonia__Effusion": 0.021505376344086023,
+    "Pneumonia__Pleural_Thickening": 0.0,
+    "Pneumonia__Cardiomegaly": 0.041666666666666664,
+    "Pneumonia__Nodule": 0.0,
+    "Pneumonia__Mass": 0.0,
+    "Pneumonia__Hernia": 0.0,
+    "Pleural_Thickening__Atelectasis": 0.04819277108433735,
+    "Pleural_Thickening__Consolidation": 0.06976744186046512,
+    "Pleural_Thickening__Infiltration": 0.02608695652173913,
+    "Pleural_Thickening__Pneumothorax": 0.0,
+    "Pleural_Thickening__Edema": 0.0,
+    "Pleural_Thickening__Emphysema": 0.037037037037037035,
+    "Pleural_Thickening__Fibrosis": 0.11627906976744186,
+    "Pleural_Thickening__Effusion": 0.0784313725490196,
+    "Pleural_Thickening__Pneumonia": 0.0,
+    "Pleural_Thickening__Cardiomegaly": 0.015625,
+    "Pleural_Thickening__Nodule": 0.04878048780487805,
+    "Pleural_Thickening__Mass": 0.13793103448275862,
+    "Pleural_Thickening__Hernia": 0.0,
+    "Cardiomegaly__Atelectasis": 0.03773584905660377,
+    "Cardiomegaly__Consolidation": 0.029850746268656716,
+    "Cardiomegaly__Infiltration": 0.09302325581395349,
+    "Cardiomegaly__Pneumothorax": 0.017543859649122806,
+    "Cardiomegaly__Edema": 0.15555555555555556,
+    "Cardiomegaly__Emphysema": 0.0,
+    "Cardiomegaly__Fibrosis": 0.028985507246376812,
+    "Cardiomegaly__Effusion": 0.17699115044247787,
+    "Cardiomegaly__Pneumonia": 0.041666666666666664,
+    "Cardiomegaly__Pleural_Thickening": 0.015625,
+    "Cardiomegaly__Nodule": 0.015384615384615385,
+    "Cardiomegaly__Mass": 0.01818181818181818,
+    "Cardiomegaly__Hernia": 0.04081632653061224,
+    "Nodule__Atelectasis": 0.0,
+    "Nodule__Consolidation": 0.09302325581395349,
+    "Nodule__Infiltration": 0.05309734513274336,
+    "Nodule__Pneumothorax": 0.02857142857142857,
+    "Nodule__Edema": 0.034482758620689655,
+    "Nodule__Emphysema": 0.0,
+    "Nodule__Fibrosis": 0.11363636363636363,
+    "Nodule__Effusion": 0.05714285714285714,
+    "Nodule__Pneumonia": 0.0,
+    "Nodule__Pleural_Thickening": 0.04878048780487805,
+    "Nodule__Cardiomegaly": 0.015384615384615385,
+    "Nodule__Mass": 0.0967741935483871,
+    "Nodule__Hernia": 0.0,
+    "Mass__Atelectasis": 0.04,
+    "Mass__Consolidation": 0.08823529411764706,
+    "Mass__Infiltration": 0.009259259259259259,
+    "Mass__Pneumothorax": 0.04,
+    "Mass__Edema": 0.0,
+    "Mass__Emphysema": 0.0,
+    "Mass__Fibrosis": 0.02631578947368421,
+    "Mass__Effusion": 0.052083333333333336,
+    "Mass__Pneumonia": 0.0,
+    "Mass__Pleural_Thickening": 0.13793103448275862,
+    "Mass__Cardiomegaly": 0.01818181818181818,
+    "Mass__Nodule": 0.0967741935483871,
+    "Mass__Hernia": 0.05555555555555555,
+    "Hernia__Atelectasis": 0.0,
+    "Hernia__Consolidation": 0.0,
+    "Hernia__Infiltration": 0.0,
+    "Hernia__Pneumothorax": 0.0,
+    "Hernia__Edema": 0.0,
+    "Hernia__Emphysema": 0.0,
+    "Hernia__Fibrosis": 0.0,
+    "Hernia__Effusion": 0.010526315789473684,
+    "Hernia__Pneumonia": 0.0,
+    "Hernia__Pleural_Thickening": 0.0,
+    "Hernia__Cardiomegaly": 0.04081632653061224,
+    "Hernia__Nodule": 0.0,
+    "Hernia__Mass": 0.05555555555555555
+  },
+  "binary_axis_top_pairs": [
+    {
+      "p1": "Pneumothorax",
+      "p2": "Fibrosis",
+      "n_ch_axis": 7,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Effusion",
+      "p2": "Hernia",
+      "n_ch_axis": 7,
+      "jaccard": 0.010526315789473684
+    },
+    {
+      "p1": "Consolidation",
+      "p2": "Hernia",
+      "n_ch_axis": 6,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Pneumothorax",
+      "p2": "Edema",
+      "n_ch_axis": 6,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Pleural_Thickening",
+      "p2": "Mass",
+      "n_ch_axis": 6,
+      "jaccard": 0.13793103448275862
+    },
+    {
+      "p1": "Emphysema",
+      "p2": "Pleural_Thickening",
+      "n_ch_axis": 5,
+      "jaccard": 0.037037037037037035
+    },
+    {
+      "p1": "Emphysema",
+      "p2": "Hernia",
+      "n_ch_axis": 5,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Fibrosis",
+      "p2": "Effusion",
+      "n_ch_axis": 5,
+      "jaccard": 0.06422018348623854
+    },
+    {
+      "p1": "Fibrosis",
+      "p2": "Hernia",
+      "n_ch_axis": 5,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Pneumothorax",
+      "p2": "Mass",
+      "n_ch_axis": 4,
+      "jaccard": 0.04
+    },
+    {
+      "p1": "Pneumothorax",
+      "p2": "Hernia",
+      "n_ch_axis": 4,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Emphysema",
+      "p2": "Cardiomegaly",
+      "n_ch_axis": 4,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Pneumonia",
+      "p2": "Hernia",
+      "n_ch_axis": 4,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Pleural_Thickening",
+      "p2": "Hernia",
+      "n_ch_axis": 4,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Mass",
+      "p2": "Hernia",
+      "n_ch_axis": 4,
+      "jaccard": 0.05555555555555555
+    },
+    {
+      "p1": "Mass",
+      "p2": "Fracture",
+      "n_ch_axis": 4,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Hernia",
+      "p2": "Enlarged Cardiomediastinum",
+      "n_ch_axis": 4,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Pneumothorax",
+      "p2": "Emphysema",
+      "n_ch_axis": 3,
+      "jaccard": 0.05
+    },
+    {
+      "p1": "Pneumothorax",
+      "p2": "Effusion",
+      "n_ch_axis": 3,
+      "jaccard": 0.0
+    },
+    {
+      "p1": "Edema",
+      "p2": "Lung Lesion",
+      "n_ch_axis": 3,
+      "jaccard": 0.0
+    }
+  ]
+}

metrics/apply_threshold_calibration.json

@@ -0,0 +1,225 @@
+{
+  "thresholds": {
+    "Atelectasis": 0.1641383171081543,
+    "Consolidation": 0.057644788175821304,
+    "Infiltration": 0.12889404594898224,
+    "Pneumothorax": 0.010074671357870102,
+    "Edema": 0.006034631747752428,
+    "Emphysema": 0.009953165426850319,
+    "Fibrosis": 0.04691587761044502,
+    "Effusion": 0.1522664874792099,
+    "Pneumonia": 0.01452414970844984,
+    "Pleural_Thickening": 0.027503035962581635,
+    "Cardiomegaly": 0.08342594653367996,
+    "Nodule": 0.04766174405813217,
+    "Mass": 0.05993369594216347,
+    "Hernia": 0.17927372455596924,
+    "Lung Lesion": 0.5,
+    "Fracture": 0.5,
+    "Lung Opacity": 0.5,
+    "Enlarged Cardiomediastinum": 0.5
+  },
+  "results": [
+    {
+      "pathology": "Atelectasis",
+      "auroc": 0.802732379979571,
+      "threshold_old": 0.5,
+      "threshold_new": 0.1641383171081543,
+      "old": {
+        "f1": 0.0851063829787234,
+        "precision": 0.6666666666666666,
+        "recall": 0.045454545454545456
+      },
+      "new": {
+        "f1": 0.4027777777777778,
+        "precision": 0.29,
+        "recall": 0.6590909090909091
+      }
+    },
+    {
+      "pathology": "Consolidation",
+      "auroc": 0.7374537227478404,
+      "threshold_old": 0.5,
+      "threshold_new": 0.057644788175821304,
+      "old": {
+        "f1": 0.0,
+        "precision": 0.0,
+        "recall": 0.0
+      },
+      "new": {
+        "f1": 0.2122905027932961,
+        "precision": 0.12418300653594772,
+        "recall": 0.7307692307692307
+      }
+    },
+    {
+      "pathology": "Infiltration",
+      "auroc": 0.6667513550135501,
+      "threshold_old": 0.5,
+      "threshold_new": 0.12889404594898224,
+      "old": {
+        "f1": 0.0,
+        "precision": 0.0,
+        "recall": 0.0
+      },
+      "new": {
+        "f1": 0.34146341463414637,
+        "precision": 0.21875,
+        "recall": 0.7777777777777778
+      }
+    },
+    {
+      "pathology": "Pneumothorax",
+      "auroc": 0.8123076923076924,
+      "threshold_old": 0.5,
+      "threshold_new": 0.010074671357870102,
+      "old": {
+        "f1": 0.0,
+        "precision": 0.0,
+        "recall": 0.0
+      },
+      "new": {
+        "f1": 0.07655502392344497,
+        "precision": 0.04020100502512563,
+        "recall": 0.8
+      }
+    },
+    {
+      "pathology": "Emphysema",
+      "auroc": 0.7564521992002907,
+      "threshold_old": 0.5,
+      "threshold_new": 0.009953165426850319,
+      "old": {
+        "f1": 0.0,
+        "precision": 0.0,
+        "recall": 0.0
+      },
+      "new": {
+        "f1": 0.0967741935483871,
+        "precision": 0.05454545454545454,
+        "recall": 0.42857142857142855
+      }
+    },
+    {
+      "pathology": "Fibrosis",
+      "auroc": 0.7666804807293826,
+      "threshold_old": 0.5,
+      "threshold_new": 0.04691587761044502,
+      "old": {
+        "f1": 0.0,
+        "precision": 0.0,
+        "recall": 0.0
+      },
+      "new": {
+        "f1": 0.1590909090909091,
+        "precision": 0.08917197452229299,
+        "recall": 0.7368421052631579
+      }
+    },
+    {
+      "pathology": "Effusion",
+      "auroc": 0.7276698577934753,
+      "threshold_old": 0.5,
+      "threshold_new": 0.1522664874792099,
+      "old": {
+        "f1": 0.04878048780487805,
+        "precision": 0.08333333333333333,
+        "recall": 0.034482758620689655
+      },
+      "new": {
+        "f1": 0.208955223880597,
+        "precision": 0.13333333333333333,
+        "recall": 0.4827586206896552
+      }
+    },
+    {
+      "pathology": "Pneumonia",
+      "auroc": 0.6679357021996615,
+      "threshold_old": 0.5,
+      "threshold_new": 0.01452414970844984,
+      "old": {
+        "f1": 0.0,
+        "precision": 0.0,
+        "recall": 0.0
+      },
+      "new": {
+        "f1": 0.056338028169014086,
+        "precision": 0.029411764705882353,
+        "recall": 0.6666666666666666
+      }
+    },
+    {
+      "pathology": "Pleural_Thickening",
+      "auroc": 0.7466666666666666,
+      "threshold_old": 0.5,
+      "threshold_new": 0.027503035962581635,
+      "old": {
+        "f1": 0.0,
+        "precision": 0.0,
+        "recall": 0.0
+      },
+      "new": {
+        "f1": 0.07272727272727272,
+        "precision": 0.0380952380952381,
+        "recall": 0.8
+      }
+    },
+    {
+      "pathology": "Cardiomegaly",
+      "auroc": 0.9929113924050632,
+      "threshold_old": 0.5,
+      "threshold_new": 0.08342594653367996,
+      "old": {
+        "f1": 0.5714285714285714,
+        "precision": 1.0,
+        "recall": 0.4
+      },
+      "new": {
+        "f1": 0.2,
+        "precision": 0.1111111111111111,
+        "recall": 1.0
+      }
+    },
+    {
+      "pathology": "Nodule",
+      "auroc": 0.5803413120567376,
+      "threshold_old": 0.5,
+      "threshold_new": 0.04766174405813217,
+      "old": {
+        "f1": 0.06896551724137931,
+        "precision": 0.2,
+        "recall": 0.041666666666666664
+      },
+      "new": {
+        "f1": 0.12807881773399016,
+        "precision": 0.07262569832402235,
+        "recall": 0.5416666666666666
+      }
+    },
+    {
+      "pathology": "Mass",
+      "auroc": 0.6666666666666666,
+      "threshold_old": 0.5,
+      "threshold_new": 0.05993369594216347,
+      "old": {
+        "f1": 0.25,
+        "precision": 0.2857142857142857,
+        "recall": 0.2222222222222222
+      },
+      "new": {
+        "f1": 0.06666666666666667,
+        "precision": 0.03482587064676617,
+        "recall": 0.7777777777777778
+      }
+    }
+  ],
+  "summary": {
+    "avg_f1_old": 0.08535674662112935,
+    "avg_f1_new": 0.16847648591212516,
+    "avg_precision_old": 0.1863095238095238,
+    "avg_precision_new": 0.10302120473709786,
+    "avg_recall_old": 0.06198551608034367,
+    "avg_recall_new": 0.7001600986061058,
+    "avg_auroc": 0.7437141189805497
+  }
+}

metrics/baseline_vs_compressed.json

@@ -0,0 +1,41 @@
+{
+  "n_test_samples": 1045,
+  "baseline_params": 6966034,
+  "compressed_params": 3383248,
+  "baseline_mean_auroc": 0.7780991651116806,
+  "compressed_mean_auroc": 0.7785241795949782,
+  "compression_pct": 51.432220973943,
+  "auroc_delta": 0.00042501448329756997,
+  "baseline_per_pathology": {
+    "Atelectasis": 0.7485004019541155,
+    "Consolidation": 0.7956338208331557,
+    "Infiltration": 0.6606524512484053,
+    "Pneumothorax": 0.7888101324657216,
+    "Edema": 0.8795579219102936,
+    "Emphysema": 0.6372302827380952,
+    "Fibrosis": 0.7325585818624467,
+    "Effusion": 0.8229508196721311,
+    "Pneumonia": 0.7992900935785737,
+    "Pleural_Thickening": 0.747797656888566,
+    "Cardiomegaly": 0.953643429202667,
+    "Nodule": 0.649755204245618,
+    "Mass": 0.7155582396014192,
+    "Hernia": 0.9614492753623188
+  },
+  "compressed_per_pathology": {
+    "Atelectasis": 0.7500051532578896,
+    "Consolidation": 0.7960815247516311,
+    "Infiltration": 0.6610364758259782,
+    "Pneumothorax": 0.7908242311565574,
+    "Edema": 0.8793965795417877,
+    "Emphysema": 0.640531994047619,
+    "Fibrosis": 0.731246195982958,
+    "Effusion": 0.8236401849516604,
+    "Pneumonia": 0.7971119716037431,
+    "Pleural_Thickening": 0.7488193624557261,
+    "Cardiomegaly": 0.953643429202667,
+    "Nodule": 0.6492693500766156,
+    "Mass": 0.7143504189627841,
+    "Hernia": 0.9633816425120774
+  }
+}

metrics/eval_nih_weights.json

@@ -0,0 +1,81 @@
+{
+  "densenet121-res224-all": {
+    "mean_auroc": 0.7780991651116805,
+    "per_pathology": {
+      "Atelectasis": 0.7485004019541155,
+      "Consolidation": 0.7956338208331557,
+      "Infiltration": 0.6606524512484053,
+      "Pneumothorax": 0.7888101324657215,
+      "Edema": 0.8795579219102936,
+      "Emphysema": 0.6372302827380952,
+      "Fibrosis": 0.7325585818624467,
+      "Effusion": 0.8229508196721311,
+      "Pneumonia": 0.7992900935785736,
+      "Pleural_Thickening": 0.747797656888566,
+      "Cardiomegaly": 0.953643429202667,
+      "Nodule": 0.649755204245618,
+      "Mass": 0.7155582396014193,
+      "Hernia": 0.9614492753623187
+    },
+    "ms_per_image": 14.936776366530424
+  },
+  "densenet121-res224-nih": {
+    "mean_auroc": 0.7524074761557032,
+    "per_pathology": {
+      "Atelectasis": 0.7468822789767693,
+      "Consolidation": 0.7790687758495716,
+      "Infiltration": 0.6114647608633392,
+      "Pneumothorax": 0.7621814238128438,
+      "Edema": 0.8793965795417877,
+      "Emphysema": 0.684477306547619,
+      "Fibrosis": 0.7376559646987219,
+      "Effusion": 0.8017822614543927,
+      "Pneumonia": 0.7049048080025814,
+      "Pleural_Thickening": 0.7584915084915085,
+      "Cardiomegaly": 0.9384542549414132,
+      "Nodule": 0.6059348955413536,
+      "Mass": 0.6944591228202612,
+      "Hernia": 0.8285507246376812
+    },
+    "ms_per_image": 15.50239398719021
+  },
+  "densenet121-res224-chex": {
+    "mean_auroc": 0.7424883832759307,
+    "per_pathology": {
+      "Atelectasis": 0.6807247541895987,
+      "Consolidation": 0.7673431970323626,
+      "Pneumothorax": 0.560965218064916,
+      "Edema": 0.8693933526944175,
+      "Effusion": 0.756754939050021,
+      "Pneumonia": 0.7333817360438851,
+      "Cardiomegaly": 0.8288554858563142
+    },
+    "ms_per_image": 15.325347088170394
+  },
+  "densenet121-res224-mimic_nb": {
+    "mean_auroc": 0.7049031491784382,
+    "per_pathology": {
+      "Atelectasis": 0.6763238719518478,
+      "Consolidation": 0.6948364814735855,
+      "Pneumothorax": 0.632446355256023,
+      "Edema": 0.8654404646660212,
+      "Effusion": 0.7896258932324506,
+      "Pneumonia": 0.46644078735075833,
+      "Cardiomegaly": 0.8092081903183809
+    },
+    "ms_per_image": 14.785645909286572
+  },
+  "densenet121-res224-mimic_ch": {
+    "mean_auroc": 0.717831448296221,
+    "per_pathology": {
+      "Atelectasis": 0.6628635623441139,
+      "Consolidation": 0.7029591097087793,
+      "Pneumothorax": 0.7114803625377644,
+      "Edema": 0.8822200709906421,
+      "Effusion": 0.7809079445145019,
+      "Pneumonia": 0.505243626976444,
+      "Cardiomegaly": 0.779145461001302
+    },
+    "ms_per_image": 14.503348729256809
+  }
+}

metrics/minimal_retrain.json

@@ -0,0 +1,131 @@
+{
+  "n_train": 3544,
+  "n_test": 1045,
+  "trainable_params": 18450,
+  "n_epochs": 30,
+  "time_forward_train": 53.12885046005249,
+  "time_forward_test": 15.548182725906372,
+  "time_classifier_train": 1.5586333274841309,
+  "time_total": 70.235666513443,
+  "best_auroc": 0.7958435729902419,
+  "per_pathology": {
+    "Atelectasis": {
+      "auroc_orig": 0.7500051532578896,
+      "auroc_new": 0.7385855337744522,
+      "f1_orig": 0.07339449541284404,
+      "f1_new": 0.01904761904761905,
+      "recall_orig": 0.038834951456310676,
+      "recall_new": 0.009708737864077669
+    },
+    "Consolidation": {
+      "auroc_orig": 0.7960815247516311,
+      "auroc_new": 0.8197458747281797,
+      "f1_orig": 0.041666666666666664,
+      "f1_new": 0.0,
+      "recall_orig": 0.02127659574468085,
+      "recall_new": 0.0
+    },
+    "Infiltration": {
+      "auroc_orig": 0.6610364758259782,
+      "auroc_new": 0.6723424197453721,
+      "f1_orig": 0.011049723756906077,
+      "f1_new": 0.07368421052631578,
+      "recall_orig": 0.005649717514124294,
+      "recall_new": 0.03954802259887006
+    },
+    "Pneumothorax": {
+      "auroc_orig": 0.7908242311565574,
+      "auroc_new": 0.8066465256797584,
+      "f1_orig": 0.0,
+      "f1_new": 0.0,
+      "recall_orig": 0.0,
+      "recall_new": 0.0
+    },
+    "Edema": {
+      "auroc_orig": 0.8793965795417877,
+      "auroc_new": 0.8959341723136496,
+      "f1_orig": 0.0,
+      "f1_new": 0.0,
+      "recall_orig": 0.0,
+      "recall_new": 0.0
+    },
+    "Emphysema": {
+      "auroc_orig": 0.640531994047619,
+      "auroc_new": 0.6788969494047619,
+      "f1_orig": 0.0,
+      "f1_new": 0.0,
+      "recall_orig": 0.0,
+      "recall_new": 0.0
+    },
+    "Fibrosis": {
+      "auroc_orig": 0.731246195982958,
+      "auroc_new": 0.7930614729153986,
+      "f1_orig": 0.0,
+      "f1_new": 0.0,
+      "recall_orig": 0.0,
+      "recall_new": 0.0
+    },
+    "Effusion": {
+      "auroc_orig": 0.8236401849516604,
+      "auroc_new": 0.8451113913408995,
+      "f1_orig": 0.30526315789473685,
+      "f1_new": 0.23952095808383234,
+      "recall_orig": 0.2230769230769231,
+      "recall_new": 0.15384615384615385
+    },
+    "Pneumonia": {
+      "auroc_orig": 0.7971119716037431,
+      "auroc_new": 0.803162310422717,
+      "f1_orig": 0.0,
+      "f1_new": 0.0,
+      "recall_orig": 0.0,
+      "recall_new": 0.0
+    },
+    "Pleural_Thickening": {
+      "auroc_orig": 0.7488193624557261,
+      "auroc_new": 0.7974298428843882,
+      "f1_orig": 0.0,
+      "f1_new": 0.0,
+      "recall_orig": 0.0,
+      "recall_new": 0.0
+    },
+    "Cardiomegaly": {
+      "auroc_orig": 0.953643429202667,
+      "auroc_new": 0.9630331005641694,
+      "f1_orig": 0.4578313253012048,
+      "f1_new": 0.49411764705882355,
+      "recall_orig": 0.37254901960784315,
+      "recall_new": 0.4117647058823529
+    },
+    "Nodule": {
+      "auroc_orig": 0.6492693500766156,
+      "auroc_new": 0.6475128003886833,
+      "f1_orig": 0.03389830508474576,
+      "f1_new": 0.0,
+      "recall_orig": 0.018518518518518517,
+      "recall_new": 0.0
+    },
+    "Mass": {
+      "auroc_orig": 0.7143504189627841,
+      "auroc_new": 0.7253717822903298,
+      "f1_orig": 0.09302325581395349,
+      "f1_new": 0.0,
+      "recall_orig": 0.07692307692307693,
+      "recall_new": 0.0
+    },
+    "Hernia": {
+      "auroc_orig": 0.9633816425120774,
+      "auroc_new": 0.9549758454106281,
+      "f1_orig": 0.7619047619047619,
+      "f1_new": 0.5333333333333333,
+      "recall_orig": 0.8,
+      "recall_new": 0.4
+    }
+  },
+  "mean_auroc_orig": 0.7785241795949782,
+  "mean_auroc_new": 0.7958435729902419,
+  "mean_f1_orig": 0.12700226370255854,
+  "mean_f1_new": 0.09712169771785173,
+  "mean_recall_orig": 0.11120205734581982,
+  "mean_recall_new": 0.0724905442993896
+}

metrics/minimal_retrain_v2.json

@@ -0,0 +1,49 @@
+{
+  "params_A": 18450,
+  "time_A": 7.310358047485352,
+  "params_B": 108309,
+  "time_B": 7.797998666763306,
+  "time_cache": 77.44846630096436,
+  "baseline_auroc": 0.7785241795949782,
+  "baseline_f1_05": 0.12700226370255854,
+  "A_thresholds": {
+    "Atelectasis": 0.07627483457326889,
+    "Consolidation": 0.013669912703335285,
+    "Infiltration": 0.1382860392332077,
+    "Pneumothorax": 0.013389605097472668,
+    "Edema": 0.006400782149285078,
+    "Emphysema": 0.01267245877534151,
+    "Fibrosis": 0.040671397000551224,
+    "Effusion": 0.18531498312950134,
+    "Pneumonia": 0.003466638969257474,
+    "Pleural_Thickening": 0.012083956971764565,
+    "Cardiomegaly": 0.00949457474052906,
+    "Nodule": 0.02881845273077488,
+    "Mass": 0.05119994655251503,
+    "Hernia": 0.5,
+    "Lung Lesion": 0.5,
+    "Fracture": 0.5,
+    "Lung Opacity": 0.5,
+    "Enlarged Cardiomediastinum": 0.5
+  },
+  "B_thresholds": {
+    "Atelectasis": 0.06119315326213837,
+    "Consolidation": 0.01689857803285122,
+    "Infiltration": 0.13535840809345245,
+    "Pneumothorax": 0.007273209281265736,
+    "Edema": 0.005829972214996815,
+    "Emphysema": 0.007729589007794857,
+    "Fibrosis": 0.0443153940141201,
+    "Effusion": 0.21586214005947113,
+    "Pneumonia": 0.00483460770919919,
+    "Pleural_Thickening": 0.013916002586483955,
+    "Cardiomegaly": 0.00925687700510025,
+    "Nodule": 0.027042903006076813,
+    "Mass": 0.07016012817621231,
+    "Hernia": 0.5,
+    "Lung Lesion": 0.5,
+    "Fracture": 0.5,
+    "Lung Opacity": 0.5,
+    "Enlarged Cardiomediastinum": 0.5
+  }
+}

metrics/q_conflict_legitimacy.json

@@ -0,0 +1,225 @@
+{
+  "polarized_ch_total": 100,
+  "legitimacy_counts": {
+    "perfect": 48,
+    "legitimate": 41,
+    "mixed": 11,
+    "conflict": 0
+  },
+  "real_conflict_ch_count": 0,
+  "real_conflict_examples": [],
+  "top_pair_legitimacy": [
+    {
+      "pair": [
+        "Fibrosis",
+        "Pneumothorax"
+      ],
+      "n_ch_polarized": 7,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Effusion",
+        "Hernia"
+      ],
+      "n_ch_polarized": 7,
+      "jaccard": 0.010526315789473684,
+      "both_positive": 1
+    },
+    {
+      "pair": [
+        "Edema",
+        "Pneumothorax"
+      ],
+      "n_ch_polarized": 6,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Mass",
+        "Pleural_Thickening"
+      ],
+      "n_ch_polarized": 6,
+      "jaccard": 0.13793103448275862,
+      "both_positive": 4
+    },
+    {
+      "pair": [
+        "Consolidation",
+        "Hernia"
+      ],
+      "n_ch_polarized": 6,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Fibrosis",
+        "Hernia"
+      ],
+      "n_ch_polarized": 5,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Effusion",
+        "Fibrosis"
+      ],
+      "n_ch_polarized": 5,
+      "jaccard": 0.06422018348623854,
+      "both_positive": 7
+    },
+    {
+      "pair": [
+        "Emphysema",
+        "Pleural_Thickening"
+      ],
+      "n_ch_polarized": 5,
+      "jaccard": 0.037037037037037035,
+      "both_positive": 1
+    },
+    {
+      "pair": [
+        "Emphysema",
+        "Hernia"
+      ],
+      "n_ch_polarized": 5,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Hernia",
+        "Pneumonia"
+      ],
+      "n_ch_polarized": 4,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Enlarged Cardiomediastinum",
+        "Hernia"
+      ],
+      "n_ch_polarized": 4,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Hernia",
+        "Pneumothorax"
+      ],
+      "n_ch_polarized": 4,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Hernia",
+        "Mass"
+      ],
+      "n_ch_polarized": 4,
+      "jaccard": 0.05555555555555555,
+      "both_positive": 1
+    },
+    {
+      "pair": [
+        "Fracture",
+        "Mass"
+      ],
+      "n_ch_polarized": 4,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Mass",
+        "Pneumothorax"
+      ],
+      "n_ch_polarized": 4,
+      "jaccard": 0.04,
+      "both_positive": 1
+    },
+    {
+      "pair": [
+        "Cardiomegaly",
+        "Emphysema"
+      ],
+      "n_ch_polarized": 4,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Hernia",
+        "Pleural_Thickening"
+      ],
+      "n_ch_polarized": 4,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Lung Opacity",
+        "Pleural_Thickening"
+      ],
+      "n_ch_polarized": 3,
+      "jaccard": 0.0,
+      "both_positive": 0
+    },
+    {
+      "pair": [
+        "Nodule",
+        "Pleural_Thickening"
+      ],
+      "n_ch_polarized": 3,
+      "jaccard": 0.04878048780487805,
+      "both_positive": 2
+    },
+    {
+      "pair": [
+        "Cardiomegaly",
+        "Hernia"
+      ],
+      "n_ch_polarized": 3,
+      "jaccard": 0.04081632653061224,
+      "both_positive": 2
+    }
+  ],
+  "pathology_jaccard_matrix": {
+    "Effusion__Cardiomegaly": 0.17699115044247787,
+    "Atelectasis__Effusion": 0.17424242424242425,
+    "Infiltration__Effusion": 0.1625,
+    "Edema__Cardiomegaly": 0.15555555555555556,
+    "Pleural_Thickening__Mass": 0.13793103448275862,
+    "Atelectasis__Infiltration": 0.13194444444444445,
+    "Fibrosis__Pleural_Thickening": 0.11627906976744186,
+    "Fibrosis__Nodule": 0.11363636363636363,
+    "Nodule__Mass": 0.0967741935483871,
+    "Atelectasis__Consolidation": 0.0963855421686747,
+    "Consolidation__Effusion": 0.09615384615384616,
+    "Consolidation__Nodule": 0.09302325581395349,
+    "Infiltration__Cardiomegaly": 0.09302325581395349,
+    "Consolidation__Mass": 0.08823529411764706,
+    "Effusion__Pleural_Thickening": 0.0784313725490196,
+    "Edema__Pneumonia": 0.07692307692307693,
+    "Consolidation__Pleural_Thickening": 0.06976744186046512,
+    "Infiltration__Fibrosis": 0.06896551724137931,
+    "Consolidation__Edema": 0.06451612903225806,
+    "Fibrosis__Effusion": 0.06422018348623854,
+    "Effusion__Nodule": 0.05714285714285714,
+    "Mass__Hernia": 0.05555555555555555,
+    "Infiltration__Nodule": 0.05309734513274336,
+    "Effusion__Mass": 0.052083333333333336,
+    "Consolidation__Infiltration": 0.05172413793103448,
+    "Pneumothorax__Emphysema": 0.05,
+    "Pleural_Thickening__Nodule": 0.04878048780487805,
+    "Atelectasis__Pleural_Thickening": 0.04819277108433735,
+    "Edema__Effusion": 0.043010752688172046,
+    "Pneumonia__Cardiomegaly": 0.041666666666666664
+  }
+}

metrics/surgery_channel_ablation.json

@@ -0,0 +1,114 @@
+{
+  "baseline": {
+    "fp_prob": 0.8913499116897583,
+    "tp_prob_avg": 0.4156641662120819,
+    "auroc": 0.9236941786283891
+  },
+  "K=1": {
+    "fp_prob": 0.8929352760314941,
+    "tp_prob_avg": 0.4644327461719513,
+    "auroc": 0.9286782296650717,
+    "tp_drop_count": 0,
+    "ablated_ch": [
+      1002
+    ]
+  },
+  "K=3": {
+    "fp_prob": 0.8070836663246155,
+    "tp_prob_avg": 0.40393105149269104,
+    "auroc": 0.9243919457735248,
+    "tp_drop_count": 0,
+    "ablated_ch": [
+      1002,
+      654,
+      961
+    ]
+  },
+  "K=5": {
+    "fp_prob": 0.7640870213508606,
+    "tp_prob_avg": 0.40278032422065735,
+    "auroc": 0.9220992822966507,
+    "tp_drop_count": 0,
+    "ablated_ch": [
+      1002,
+      654,
+      961,
+      1012,
+      665
+    ]
+  },
+  "K=10": {
+    "fp_prob": 0.6739818453788757,
+    "tp_prob_avg": 0.41819897294044495,
+    "auroc": 0.927731259968102,
+    "tp_drop_count": 4,
+    "ablated_ch": [
+      1002,
+      654,
+      961,
+      1012,
+      665,
+      964,
+      1005,
+      657,
+      644,
+      973
+    ]
+  },
+  "K=20": {
+    "fp_prob": 0.5440241098403931,
+    "tp_prob_avg": 0.43110406398773193,
+    "auroc": 0.9224980063795853,
+    "tp_drop_count": 7,
+    "ablated_ch": [
+      1002,
+      654,
+      961,
+      1012,
+      665,
+      964,
+      1005,
+      657,
+      644,
+      973,
+      978,
+      1003,
+      980,
+      1021,
+      110,
+      983,
+      899,
+      1000,
+      757,
+      590
+    ]
+  },
+  "K=30": {
+    "fp_prob": 0.5440241098403931,
+    "tp_prob_avg": 0.43110406398773193,
+    "auroc": 0.9224980063795853,
+    "tp_drop_count": 7,
+    "ablated_ch": [
+      1002,
+      654,
+      961,
+      1012,
+      665,
+      964,
+      1005,
+      657,
+      644,
+      973,
+      978,
+      1003,
+      980,
+      1021,
+      110,
+      983,
+      899,
+      1000,
+      757,
+      590
+    ]
+  }
+}

requirements.txt

@@ -0,0 +1,6 @@
+torch>=2.0
+torchvision
+torchxrayvision>=1.0
+pillow
+numpy
+scikit-learn