README.md

---
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

Zenodo preprint: [10.5281/zenodo.20131680](https://doi.org/10.5281/zenodo.20131680)

```bibtex
@misc{cynn2026chexnet,
  title={Lossless Mechanistic Compression and Surgical Correction of Medical Imaging Models},
  author={Cynn, Yeonseong},
  year={2026},
  publisher={Zenodo},
  doi={10.5281/zenodo.20131680},
  url={https://doi.org/10.5281/zenodo.20131680}
}
```

## 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