🩺 Thyroid Nodule AI: Deep Learning CAD Pipeline

Computer-Aided Diagnosis (CAD) system for Thyroid Ultrasound Analysis.

This repository contains the official implementation of the Bachelor's Thesis: "Development of a Deep Learning pipeline for thyroid nodule diagnosis: Comparison between CNN and Vision Transformers" (Sapienza University of Rome, 2025).

---

📖 Table of Contents

• Project Overview
• Methodology & Pipeline
• Key Results
• Clinical Integration & Risk Stratification
• Installation
• Usage (GUI & CLI)
• Repository Structure
• Citation

---

🏥 Project Overview

Thyroid nodules are a pervasive clinical issue, present in up to 60% of the adult population. While the majority are benign, distinguishing the 5-10% of malignant cases remains a challenge due to the subjective nature of ultrasound interpretation (high inter-observer variability).

This project proposes a two-stage Deep Learning pipeline to support radiologists:

1. Detection: Localizing nodules in B-mode ultrasound images.
2. Classification: Discriminating between Benign and Malignant nodules.
3. Explainability: Visualizing the morphological features driving the AI's decision.

Research Goal

The core study compares traditional CNNs (EfficientNet, ConvNeXt) against modern Vision Transformers (DINOv3, Swin) to determine if Self-Supervised Learning (SSL) offers superior robustness in processing noisy medical imagery.

---

⚙️ Methodology & Pipeline

The system processes raw ultrasound images through a strict pipeline described in Chapter 3 of the thesis.

1. Preprocessing & Enhancement

Ultrasound images suffer from speckle noise and low contrast. Before inference, images undergo:

• Perceptual Hashing (dHash): To remove duplicate frames and prevent data leakage.
• CLAHE: Contrast Limited Adaptive Histogram Equalization.
• Sharpening: To emphasize nodule margins (a critical TI-RADS feature).

Figure: Effect of the enhancement pipeline. (Left) Original raw input. (Right) Enhanced image fed to the models.

2. Architecture Comparison

We benchmarked the following architectures:

• Detection: YOLOv12 (Anchor-Free) vs. DINO-DETR vs. Faster R-CNN.
• Classification:
  • Baselines: SVM (Radiomics), EfficientNetV2, ConvNeXt V2.
  • Our Approach: DINOv3 (ViT) pretrained with Self-Supervised Learning (SSL).

---

📊 Key Results

The models were evaluated on the TN5000 and AUITD datasets (7,000+ nodules).

🏆 Classification Performance (Test Set)

DINOv3-Large achieved state-of-the-art results, significantly outperforming CNNs in Sensitivity (Recall), which is the most critical metric for cancer screening.

Architecture	Type	AUC-ROC	Recall (Sensitivity)	F1-Score
EfficientNetV2	CNN	0.898	92.63%	0.864
ConvNeXt V2	CNN	0.906	91.71%	0.865
DINOv3-Large	ViT	0.932	94.70%	0.887

Visual Analysis

• Left (ROC Curve): The Foundation Model (Orange line) demonstrates superior separation capabilities ($AUC=0.93$).
• Right (Confusion Matrix): High accuracy in identifying malignant cases, minimizing dangerous False Negatives.

---

🩺 Clinical Integration & Risk Stratification

The system is designed as a Decision Support Tool. While the model performs a binary classification (Benign vs. Malignant) using an optimized operational threshold ($p=0.38$), the raw probability score ($p$) provides a granular estimation of malignancy risk.

Validation on external datasets (Chapter 5.5.4) confirms that the model's confidence levels correlate strongly with the K-TIRADS risk categories, allowing for a professional interpretation of the AI output:

AI Confidence ($p$)	Classification	Clinical Interpretation	Suggested Action
$p < 0.20$	Benign	Low risk (Correlates to TR1/TR2)	Routine follow-up
$0.20 \le p < 0.60$	Indeterminate	Moderate risk (Correlates to TR3/TR4)	Short-term follow-up / FNA
$p \ge 0.60$	Malignant	High Risk (Correlates to TR5)	Strong Biopsy Recommendation

🔍 Explainability (Heatmaps)

To move beyond "Black Box" AI, the pipeline integrates Attention Maps (for DINOv3) and Grad-CAM (for CNNs). This allows clinicians to verify if the AI is focusing on relevant radiological features, such as:

• Irregular Margins: Precisely outlined by the Transformer's global attention.
• Microcalcifications: Detected as high-frequency textures by the CNN backbones.

Figure: DINOv3 attention maps focusing on the irregular borders of a malignant nodule, aligning with K-TIRADS criteria.

---

💻 Installation

Prerequisites

• OS: Linux (Recommended) or Windows 10/11
• Python: 3.10+
• Hardware: NVIDIA GPU with CUDA support (Recommended for DINOv3)

Setup Steps

# 1. Clone the repository
git clone https://github.com/YourUsername/Thyroid-Nodule-AI.git
cd Thyroid-Nodule-AI

# 2. Create a virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# 3. Install core dependencies (PyTorch, Ultralytics, etc.)
pip install -r requirements.txt

# 4. (Optional) Install GUI dependencies
pip install -r requirements-gui.txt

⚠️ Model Weights: Due to file size limits and licensing, weights are not included in the repo. Please refer to docs/WEIGHTS.md for download links and placement instructions.

---

🚀 Usage (GUI & CLI)

1. Thyroid AI Assistant (GUI)

A user-friendly interface for simulating the clinical workflow (Drag & Drop, Real-time Analysis).

python src/gui/app.py

Features:

• Visual toggle for Preprocessing (CLAHE).
• Real-time Detection (YOLO) and Classification (DINO).
• PDF Report generation (Experimental).

2. Command Line Interface (CLI)

Run the classifier on a single image via terminal.

python3 src/inference_dino.py \
  --image path/to/your/image.jpg \
  --weights path/to/model_weights.pt \
  --dino-repo dinov3 \
  --out-dir results

Arguments:

Argument	Description	Example
--image	Path to the input image (supports .jpg, .png).	assets/benign.jpg
--weights	Path to the trained DINOv3 .pt file.	models/dinov3_large.pt
--dino-repo	Path to the local DINOv3 repository/folder.	dinov3
--out-dir	Directory where the result will be saved.	results

Note: The script automatically applies the preprocessing pipeline (CLAHE + Sharpening) before inference.

---

📂 Repository Structure

.
├── assets/                 # Images, plots, and demo resources
│   ├── benign              # Images of benign nodules
│   └── malignant           # Images of malignant nodules
├── dinov3/                 # Submodule for Vision Transformer architecture
├── docs/                   # Documentation
│   ├── GUI.md              # User manual for the interface
│   └── WEIGHTS.md          # Links to pretrained models
├── src/                    # Source code
│   ├── gui/                # Application logic (CustomTkinter)
│   └── inference_dino.py   # CLI inference script
├── Thesis.pdf              # Full Bachelor's Thesis document
├── CITATION.bib            # BibTeX citation
├── LICENSE                 # MIT License
├── MODEL_CARD.md           # Model specifics and limitations
├── requirements.txt        # Python dependencies
└── requirements-gui.txt    # gui dependencies

---

🎓 Citation

If you use this work in your research, please cite the thesis:

@bachelorsthesis{catania2025thyroid,
  author  = {Alessandro Catania},
  title   = {Development of a Deep Learning pipeline for thyroid nodule diagnosis: Comparison between CNN and Vision Transformers},
  school  = {Sapienza University of Rome},
  year    = {2025},
  type    = {Bachelor's Thesis}
}

---

Acknowledgments

• Sapienza University of Rome - Faculty of Information Engineering.
• Ultralytics for the YOLOv12 framework.
• Meta AI for the DINOv3 foundation model.

---

For questions or collaborations, please open an Issue.