ahmetcangunay/MammoTagger_encoderfile

🩺 MammoTagger: Turkish Mammography NER (.encoderfile)

This repository contains a standalone, zero-dependency executable binary (.encoderfile) built with Mozilla AI's Encoderfile specification. Fine-tuned from ahmetcangunay/MammoTagger, it is designed for Named Entity Recognition (NER) on Turkish mammography reports.

🏆 TEKNOFEST Achievement: Developed as part of the project that achieved 5th Place & Finalist status in the TEKNOFEST 2024 Sağlıkta Yapay Zekâ Yarışması (Üniversite ve Üzeri Seviyesi - Bilgisayarlı Görüyle Hastalık Tespiti Kategorisi).

---

🚀 Quick Start / Instant CLI Usage

You can download and run this executable directly on any x86_64 Linux machine without installing Python, PyTorch, Transformers, or SpaCy.

1. Download & Prepare Binary

# Download the executable via Hugging Face CLI
hf download ahmetcangunay/MammoTagger_encoderfile mammo-tagger.x86_64-unknown-linux-gnu.encoderfile --local-dir .

# Make it executable
chmod +x mammo-tagger.x86_64-unknown-linux-gnu.encoderfile

2. Direct CLI Token Inference (infer)

./mammo-tagger.x86_64-unknown-linux-gnu.encoderfile infer "BILATERAL MAMOGRAFI INCELEMESI: Sağ meme üst dış kadranda yaklaşık 1 cm çapında düzgün sınırlı nodüler lezyon izlendi."

Example Response:

{
  "results": [
    {
      "tokens": [
        {
          "token_info": { "token": "Sağ", "token_id": 3644, "start": 32, "end": 36 },
          "label": "B-ANAT",
          "score": 7.4060626
        },
        {
          "token_info": { "token": "meme", "token_id": 13135, "start": 37, "end": 41 },
          "label": "I-ANAT",
          "score": 7.988982
        },
        {
          "token_info": { "token": "yaklaşık", "token_id": 3870, "start": 62, "end": 72 },
          "label": "B-OBS-PRESENT",
          "score": 6.614647
        },
        {
          "token_info": { "token": "1", "token_id": 21, "start": 73, "end": 74 },
          "label": "I-OBS-PRESENT",
          "score": 6.538508
        }
      ]
    }
  ],
  "model_id": "mammo-tagger",
  "metadata": {}
}

🌐 Serving Options

REST API Server (serve)

Start a lightweight local REST server:

./mammo-tagger.x86_64-unknown-linux-gnu.encoderfile serve --http-port 8080

Anthropic MCP Server (mcp)

Start as a Model Context Protocol (MCP) server for integration with LLM agents:

./mammo-tagger.x86_64-unknown-linux-gnu.encoderfile mcp

📊 Model Performance & Metrics

Evaluated on an independent test dataset of 257 clinical mammography report sentences (18,893 total evaluated tokens):

• Overall Accuracy: 93.14%

• Weighted F1-Score: 0.9308

• Macro Average F1-Score: 0.9254

Category	Precision	Recall	F1-Score	Support
ANAT	0.9479	0.9919	0.9694	5416
IMPRESSION	0.9960	0.9861	0.9910	502
O	0.9466	0.8684	0.9058	6268
OBS-ABSENT	0.9297	0.9627	0.9460	2227
OBS-PRESENT	0.8868	0.9277	0.9068	4368
OBS-UNCERTAIN	0.8654	0.8036	0.8333	112

🏷️ Entity Types

• ANAT: Anatomical regions (e.g., sağ meme, üst dış kadran)
• IMPRESSION: Overall clinical conclusion / summary notes
• OBS-PRESENT: Present findings (e.g., nodül, kitle, mikrokalsifikasyon)
• OBS-ABSENT: Negative findings / Absence of findings (e.g., kitle saptanmadı)
• OBS-UNCERTAIN: Doubtful / Suspicious findings (e.g., şüpheli görünüm)
• O: Outside / Non-entity tokens

⚙️ Technical Details

• Base Architecture: ahmetcangunay/MammoTagger (Fine-tuned from akdeniz27/bert-base-turkish-cased-ner)
• Runtime: Standalone CPU Executable (Rust/C++ Bindings via Mozilla Encoderfile)
• Tagging Scheme: IOB2 (B-, I-, O)
• Target OS: x86_64-unknown-linux-gnu