Cancer Pathos

Cancer_pathos — Histopathologic Cancer Detection (ConvNeXt-Tiny)

Binary classification of metastatic cancer in 96×96 histopathologic tissue patches (PatchCamelyon / Kaggle Histopathologic Cancer Detection). Backbone ConvNeXt-Tiny

• a custom residual classification head (1 logit, sigmoid).

⚠️ Research Use Only (RUO) — this is not a certified medical device and must not be used for clinical diagnosis.

🔗 Code & full report: https://github.com/Hakim78/histopathologic-cancer-detection 🤗 Interactive demo: https://huggingface.co/spaces/hakim78/Cancer_pathos_demo

---

📦 Checkpoints

File	Training data	Role	Headline metric
models/best_model_epoch50_auc0.9980.pth	~95 % of labels, 50 epochs	v1 — deployed in the demo	Validation AUC 0.9983 (TTA ×5)
models/final_model_v2_locked.pth	70 % train, early-stopped on val	v2 — rigorous evaluation	Locked-test AUC 0.9979 (TTA ×5)

• v1 maximizes performance (more data, longer training) and powers the live demo.
• v2 was trained under a stricter protocol — a 3-way split with a strictly locked test set + 5-fold cross-validation — to report a defensible, independent test metric.

---

📊 Results (v2 — 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 — opened once for the final metrics. Robustness first estimated by 5-fold cross-validation on train+val.

Metric	5-fold CV (mean ± σ)	Locked test (TTA ×5)
AUC-ROC	0.9975 ± 0.0001	0.9979
Accuracy	0.9821 ± 0.0008	0.9847
F1-score	0.9778 ± 0.0010	0.9810
Brier Score	0.0212 ± 0.0005	0.0172
Sensitivity	0.9753 ± 0.0015	0.9794
Specificity	0.9867 ± 0.0006	0.9882

Calibration on the locked test: Brier 0.0172, ECE 0.0449. Confusion matrix (33,004 patches): TN 19,405 · FP 231 · FN 275 · TP 13,093. TTA ×5 yields a small but statistically significant gain (DeLong on AUC: z = 6.62, p < 10−4). Full metrics in results_v2/.

---

🏗 Architecture & training

• Backbone: timm convnext_tiny (ImageNet-pretrained, num_classes=0, drop_rate=0.3) → 768-d features.
• Head: Residual MLP — Linear(768→256) → BN → GELU → Dropout → Linear(256→256) → BN → (+ shortcut) → GELU → Dropout → Linear(256→1).
• Loss: Focal Loss (α = 0.25, γ = 2.0). Optimizer: AdamW (lr 1e-4, wd 1e-5). Scheduler: CosineAnnealingLR. AMP enabled. Seed: 42.
• Input: RGB 96×96, ImageNet normalization (mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225]).

---

🚀 Usage

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) — 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) — 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 + 90° rotation + transpose).

---

⚠️ Limitations

• Patch-level only (96×96), not whole-slide. A 1.18 % per-patch false-positive rate aggregates over the tens of thousands of patches in a slide → a MIL aggregation step is required for slide-level use.
• PatchCamelyon originates from a limited set of centers/scanners → external multicentric validation and stain normalization are required before any clinical perspective.
• Calibration is good but shows a slight sigmoid profile (ECE 0.045) → temperature scaling recommended before any autonomous use.

---

👤 Author

Hakim Djaalal — Master's in Big Data & AI. License: MIT (Research Use Only).