Dental-Implant-Classification

This repository contains the official implementation for the dental implant classification model using the AI-Hub public dataset. For further requests regarding the dataset, source code, or model weights, please contact the authors at sangyeonlee@catholic.ac.kr

Project Overview

• Objective: Open-set classification and specification estimation of dental implants from radiographic images.
• Dataset: AI-Hub Radiographic Image Data of Dental Implants
• Framework: PyTorch

Deep learning models for dental implant identification are usually trained and evaluated on a fixed set of implant systems. In practice, however, new systems keep entering the market, and a closed-set model silently assigns every unregistered implant to a known system, often with high confidence.

This repository contains the code for our study, which evaluates implant identification under open-set conditions: the model must identify registered systems and flag implants from systems it has never seen.

Code repository

• common.py Shared dataset, image transforms, and backbone utilities.
• arpl.py ARPL loss following the official implementation.
• openmax.py OpenMax and MAV distance scoring on classifier logits.
• osr_metrics.py Open-set evaluation metrics, including AUROC, OSCR, and threshold-based scores.
• train_ce.py Trains the cross-entropy classifier and saves logits, features, and ODIN scores.
• train_arpl.py Trains the ARPL model and saves its outputs and evaluation metrics.
• eval_baselines.py Compares all open-set methods on identical splits with paired statistical tests.
• check_openmax.py Runs OpenMax sanity checks and hyperparameter sensitivity analysis.
• analyze_all.py Aggregates subgroup and error analyses over all splits into paper-ready tables.
• aggregate.py Summarizes per-split metrics as mean ± standard deviation. All experiments were run on Google Colab with a single NVIDIA Tesla T4 GPU. Training takes about one hour per split; all evaluation and figure scripts run on CPU.

Logits

Sample logit outputs from the unknown test set and validation set are provided for reference.