| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - image-classification |
| - medical-imaging |
| - histopathology |
| - breast-cancer |
| - pytorch |
| - convnext |
| - patchcamelyon |
| pipeline_tag: image-classification |
| --- |
| |
| # Cancer_pathos β Histopathologic Cancer Detection (ConvNeXt-Tiny) |
| |
| Binary classification of **metastatic cancer in 96Γ96 histopathologic tissue patches** |
| (PatchCamelyon / Kaggle *Histopathologic Cancer Detection*). Backbone **ConvNeXt-Tiny** |
| + a custom **residual classification head** (1 logit, sigmoid). |
| |
| > β οΈ **Research Use Only (RUO)** β this is **not** a certified medical device and must **not** be used for clinical diagnosis. |
| |
| π **Code & full report:** https://github.com/Hakim78/histopathologic-cancer-detection |
| π€ **Interactive demo:** https://huggingface.co/spaces/hakim78/Cancer_pathos_demo |
| |
| --- |
| |
| ## π¦ Checkpoints |
| |
| | File | Training data | Role | Headline metric | |
| |------|---------------|------|-----------------| |
| | `models/best_model_epoch50_auc0.9980.pth` | ~95 % of labels, 50 epochs | **v1 β deployed in the demo** | Validation AUC **0.9983** (TTA Γ5) | |
| | `models/final_model_v2_locked.pth` | 70 % train, early-stopped on val | **v2 β rigorous evaluation** | Locked-test AUC **0.9979** (TTA Γ5) | |
|
|
| - **v1** maximizes performance (more data, longer training) and powers the live demo. |
| - **v2** was trained under a stricter protocol β a 3-way split with a **strictly locked test set** + 5-fold cross-validation β to report a **defensible, independent** test metric. |
|
|
| --- |
|
|
| ## π Results (v2 β rigorous protocol) |
|
|
| 70 / 15 / 15 stratified split (seed 42). The 15 % **test set (33,004 patches) is locked**: |
| never used for training, model selection or hyperparameters β opened once for the final metrics. |
| Robustness first estimated by 5-fold cross-validation on train+val. |
|
|
| | Metric | 5-fold CV (mean Β± Ο) | Locked test (TTA Γ5) | |
| |--------|:--------------------:|:--------------------:| |
| | AUC-ROC | 0.9975 Β± 0.0001 | **0.9979** | |
| | Accuracy | 0.9821 Β± 0.0008 | **0.9847** | |
| | F1-score | 0.9778 Β± 0.0010 | **0.9810** | |
| | Brier Score | 0.0212 Β± 0.0005 | **0.0172** | |
| | Sensitivity | 0.9753 Β± 0.0015 | **0.9794** | |
| | Specificity | 0.9867 Β± 0.0006 | **0.9882** | |
|
|
| Calibration on the locked test: **Brier 0.0172, ECE 0.0449**. |
| Confusion matrix (33,004 patches): **TN 19,405 Β· FP 231 Β· FN 275 Β· TP 13,093**. |
| TTA Γ5 yields a small but statistically significant gain (DeLong on AUC: z = 6.62, p < 10β»β΄). |
| Full metrics in [`results_v2/`](results_v2). |
|
|
| --- |
|
|
| ## π Architecture & training |
|
|
| - **Backbone:** `timm` `convnext_tiny` (ImageNet-pretrained, `num_classes=0`, `drop_rate=0.3`) β 768-d features. |
| - **Head:** Residual MLP β `Linear(768β256) β BN β GELU β Dropout β Linear(256β256) β BN β (+ shortcut) β GELU β Dropout β Linear(256β1)`. |
| - **Loss:** Focal Loss (Ξ± = 0.25, Ξ³ = 2.0). **Optimizer:** AdamW (lr 1e-4, wd 1e-5). **Scheduler:** CosineAnnealingLR. **AMP** enabled. **Seed:** 42. |
| - **Input:** RGB 96Γ96, ImageNet normalization (mean `[0.485, 0.456, 0.406]`, std `[0.229, 0.224, 0.225]`). |
|
|
| --- |
|
|
| ## π Usage |
|
|
| ```python |
| import torch, timm |
| import torch.nn as nn |
| from huggingface_hub import hf_hub_download |
| |
| class ResidualHead(nn.Module): |
| def __init__(self, in_features, hidden=256, dropout=0.3): |
| super().__init__() |
| self.fc1, self.bn1 = nn.Linear(in_features, hidden), nn.BatchNorm1d(hidden) |
| self.fc2, self.bn2 = nn.Linear(hidden, hidden), nn.BatchNorm1d(hidden) |
| self.fc_out, self.shortcut = nn.Linear(hidden, 1), nn.Linear(in_features, hidden) |
| self.gelu, self.dropout = nn.GELU(), nn.Dropout(dropout) |
| def forward(self, x): |
| identity = self.shortcut(x) |
| out = self.dropout(self.gelu(self.bn1(self.fc1(x)))) |
| out = self.gelu(self.bn2(self.fc2(out)) + identity) |
| return self.fc_out(self.dropout(out)) |
| |
| class CancerClassifier(nn.Module): |
| def __init__(self): |
| super().__init__() |
| self.backbone = timm.create_model("convnext_tiny", pretrained=False, num_classes=0) |
| self.head = ResidualHead(self.backbone.num_features) |
| def forward(self, x): |
| return self.head(self.backbone(x)) |
| |
| model = CancerClassifier() |
| |
| # v2 (rigorous) β raw state_dict: |
| path = hf_hub_download("hakim78/Cancer_pathos", "models/final_model_v2_locked.pth") |
| model.load_state_dict(torch.load(path, map_location="cpu")) |
| |
| # v1 (deployed) β checkpoint dict: |
| # path = hf_hub_download("hakim78/Cancer_pathos", "models/best_model_epoch50_auc0.9980.pth") |
| # model.load_state_dict(torch.load(path, map_location="cpu")["model_state_dict"]) |
| |
| model.eval() |
| ``` |
|
|
| Prediction = `torch.sigmoid(model(x))` (probability of cancer; threshold 0.5). For best results, average over 5 TTA views (original + H/V flip + 90Β° rotation + transpose). |
|
|
| --- |
|
|
| ## β οΈ Limitations |
|
|
| - **Patch-level only** (96Γ96), not whole-slide. A 1.18 % per-patch false-positive rate aggregates over the tens of thousands of patches in a slide β a **MIL aggregation** step is required for slide-level use. |
| - PatchCamelyon originates from a **limited set of centers/scanners** β external multicentric validation and stain normalization are required before any clinical perspective. |
| - Calibration is good but shows a slight sigmoid profile (ECE 0.045) β **temperature scaling** recommended before any autonomous use. |
|
|
| --- |
|
|
| ## π€ Author |
|
|
| **Hakim Djaalal** β Master's in Big Data & AI. License: MIT (Research Use Only). |
|
|
