language:
- en
license: cc-by-4.0
tags:
- medical
- histopathology
- cancer-detection
- binary-classification
- efficientnet
- pytorch
- onnx
- int8
metrics:
- accuracy
- recall
- f1
- auc
library_name: timm
pipeline_tag: image-classification
model-index:
- name: Aegis Stage 1 Binary Cancer Detector
results:
- task:
type: image-classification
name: Histopathology Cancer vs Benign Detection
metrics:
- name: Test Accuracy
type: accuracy
value: 0.9209
- name: Test Recall (Sensitivity)
type: recall
value: 0.9456
- name: Test Precision
type: precision
value: 0.8821
- name: Test F1-Score
type: f1
value: 0.9127
- name: ROC-AUC
type: auc
value: 0.9814
Aegis Stage 1 — Binary Cancer Detector (EfficientNet-B0)
Aegis Stage 1 is a lightweight, high-sensitivity binary classifier trained on 30,453 histopathology tiles across 6 public datasets to distinguish cancerous from non-cancerous tissue. It serves as the first stage in the Aegis cascaded pipeline for automated cancer detection and OncoTree subtype classification.
- Architecture: EfficientNet-B0 (
timm) - Total Parameters: 4,008,829 (~4.01 Million)
- INT8 ONNX Model Size: 15.7 MB
- Test Accuracy: 92.09%
- Test Recall (Cancer Sensitivity): 94.56%
- Test F1-Score: 91.27%
- ROC-AUC: 0.9814
- License: Creative Commons Attribution 4.0 (CC-BY-4.0)
Performance Benchmarks (Unseen Test Set)
Evaluated across 3,528 unseen histopathology test tiles (6 datasets):
| Dataset | Accuracy | Recall | Precision | F1-Score | ROC-AUC | Missed Cancer Rate |
|---|---|---|---|---|---|---|
| LC25000 | 99.34% | 98.92% | 100.00% | 99.46% | 0.9979 | 1.08% |
| Camelyon17 (djghosh) | 98.75% | 98.22% | 99.28% | 98.75% | 0.9987 | 1.78% |
| Camelyon17 (jxie) | 96.12% | 96.17% | 93.39% | 94.76% | 0.9960 | 3.83% |
| Patch Camelyon | 88.83% | 93.59% | 85.39% | 89.30% | 0.9656 | 6.41% |
| Breast Histopathology | 87.08% | 82.01% | 76.73% | 79.28% | 0.9371 | 17.99% |
| Skin Lesion HM10000 | 81.03% | 92.55% | 67.18% | 77.85% | 0.9201 | 7.45% |
| Overall (Weighted) | 92.09% | 94.56% | 88.21% | 91.27% | 0.9814 | 5.44% |
High-Recall Operating Mode (Clinical Screening)
At threshold 0.32, the model achieves 95.07% cancer recall with 89.32% specificity — ideal for triage/screening workflows where missed cancers must be minimized.
Dataset Attribution & Citation
This model was trained on public histopathology research datasets:
- LC25000 (CC-BY-4.0) — Lung & Colon (25,000 patches)
- Patch Camelyon (CC0-1.0) — Lymph node metastases
- Breast Histopathology Patches (CC-BY-4.0) — IDC breast cancer
- Skin Lesion HM10000 (MIT) — Melanoma/benign
- Camelyon17 (jxie/djghosh) — Lymph node WSI patches (2 subsets)
Medical Research Disclaimer
Aegis is an open-source AI research prototype intended for educational, scientific evaluation, and research purposes only. It is not a certified medical device and must not be used for primary clinical diagnosis or treatment planning.
Related Models
- Stage 2 OncoTree Subtype Classifier:
AegisOSS/stage-2— 9-class ResNet-50 for cancer subtyping
Citation
@software{aegis_stage1,
title = {Aegis Stage 1: EfficientNet-B0 Binary Cancer Detector for Histopathology},
author = {Ranveer Soni},
year = {2026},
url = {https://huggingface.co/AegisOSS/stage-1},
license = {CC-BY-4.0}
}