DentalGemma is a domain-adapted extension of MedGemma 1.5 4B IT specialized for dental diagnostics and structured clinical reasoning. The model leverages targeted multimodal fine-tuning to enable detailed interpretation of dental imaging and comprehensive oral pathology assessment.
Built for the MedGemma Impact Challenge, DentalGemma demonstrates how medical foundation models can be efficiently adapted to underserved healthcare domains using parameter-efficient fine-tuning.
🎯 Model Description
DentalGemma is a multimodal vision-language model that combines:
Visual understanding of dental images (clinical photographs and radiographs)
Clinical reasoning for diagnosis, treatment planning, and patient counseling
Structured assessment following evidence-based dental protocols
The model was fine-tuned using LoRA (Low-Rank Adaptation) in a two-stage training pipeline. Unlike typical QLoRA approaches, this model was trained in full bfloat16 precision on an NVIDIA A100 GPU to maximize diagnostic accuracy and avoid quantization artifacts.
Stage 1 (VQA): Multimodal training on 2,529 dental X-ray image-text pairs
Stage 2 (Instruct): Text-only training on 2,494 clinical case assessments
Key Capabilities
Capability
Description
📸 Clinical Photo Analysis
Analyze clinical dental photographs for cavity detection, oral health assessment, and severity evaluation with compositionally-varied clinical descriptions
🏥 Pathology Classification
Classify 6 dental conditions from panoramic X-rays (Healthy, Caries, Impacted Teeth, BDC-BDR, Infection, Fractured Teeth) with differential diagnosis and urgency assessment
📍 Location-Aware Diagnosis
Identify and localize pathological findings in panoramic radiographs using dental region mapping (e.g., "right mandibular region", "anterior maxillary region")
🦷 Dentition Assessment
Evaluate dentition completeness, tooth type identification, and anatomical overview from panoramic radiographs with clinical context
📋 Structured Radiographic Reports
Generate systematic dental reports with region-specific findings, differential diagnoses, and clinical recommendations
💊 Clinical Case Analysis
Comprehensive diagnosis, treatment planning, antibiotic considerations, and follow-up scheduling for 98 dental conditions
Format: Dental images (clinical photographs and radiographs) paired with diverse clinical questions and compositionally-generated expert answers
Sources:
Clinical Photo Analysis (~642 pairs from 418 images): Clinical dental photographs with YOLO-OBB annotations for cavity/normal regions. Generates 1-2 questions per image across 5 question types (binary classification, clinical description, severity assessment, image type identification, treatment recommendations).
OPG Classification (~1,214 pairs from 517 images): Panoramic radiographs in 6 pathology classes (Healthy Teeth, Caries, Impacted teeth, BDC-BDR, Infection, Fractured Teeth). Generates 2-3 questions per image across 5 question types (open-ended diagnosis, yes/no pathology screening, differential diagnosis, clinical urgency, healthy vs abnormal).
Panoramic Dental X-ray (~128 pairs from 64 images): Panoramic radiographs with VIA polygon annotations (tooth segmentation) and COCO annotations (8 tooth type classes). Generates 2 questions per image focusing on dentition completeness, anatomical overview, and tooth type identification.
OPG Object Detection (~545 pairs from 232 images): Panoramic radiographs with YOLO bounding box annotations for 6 pathology classes. Generates 2-3 location-aware questions per image by converting normalized bounding box coordinates to dental region descriptions (e.g., "right mandibular region", "anterior maxillary region"). Question types include localized findings, condition presence screening, structured radiographic reports, and region-specific queries.
Case Structure: Each case includes patient demographics, chief complaint, clinical findings, radiographic findings, medical history, and a comprehensive structured assessment with diagnosis, management plan, antibiotic considerations, follow-up recommendations, and patient counseling
All data follows a consistent chat template format (system/user/assistant) with a standardized system prompt emphasizing evidence-based assessment and clinical correlation.
🚀 Training Procedure
Fine-Tuning Method: Full bfloat16 LoRA
To ensure the highest fidelity in X-ray analysis, we trained the model in native bfloat16 precision without quantization:
Base Model: Loaded in full bfloat16 (no 4-bit quantization)
Adapters: LoRA (Rank 64, Alpha 64) trained on all linear modules
Hardware: NVIDIA A100 (80GB VRAM)
Training Hyperparameters
Stage 1: VQA Training (Multimodal)
Parameter
Value
Epochs
5 (Stopped early at ~4 due to convergence)
Batch Size
1 per device × 4 gradient accumulation
Learning Rate
5e-5 (linear scheduler)
Warmup Ratio
0.1
Max Sequence Length
1024
Optimizer
AdamW (fused)
Precision
bfloat16 (Full)
LoRA Rank
64
LoRA Alpha
64
LoRA Dropout
0.05
Target Modules
All linear layers
Max Grad Norm
1.0
Training Time
~4 hours
Stage 2: Instruct Training (Text-only)
Parameter
Value
Epochs
5 (Stopped at ~4 due to convergence)
Batch Size
2 per device × 4 gradient accumulation
Learning Rate
5e-5 (linear scheduler)
Warmup Ratio
0.1
Max Sequence Length
1024
Max Grad Norm
1.0
Training Results
VQA Training (Stage 1)
Step
Training Loss
Validation Loss
100
1.5756
1.3255
500
0.1846
0.1917
1000
0.1611
0.1665
1500
0.1408
0.1628
2000
0.1466
0.1612
2100
0.1339
0.1585 (Best)
2300
0.1150
0.1595
Note: Training was stopped early after Step 2300 as validation loss began to plateau. The checkpoint from Step 2100 was selected as the optimal VQA model.
Final VQA Metrics:
Best Validation Loss: 0.1585
Improvement: Significant reduction from initial loss of ~1.32
Instruct Training (Stage 2)
Step
Training Loss
Validation Loss
100
0.2984
0.2111
200
0.0594
0.0447
300
0.0131
0.0331
400
0.0100
0.0293
500
0.0045
0.0224 (Best)
600
0.0068
0.0225
1000
0.0018
0.0270
Note: Training was stopped early after Step 1000 as validation loss began to rise. The checkpoint from Step 500 was selected as the final model.
Final Instruct Metrics:
Best Validation Loss: 0.0224 (at step 500)
Training Loss at Best Checkpoint: 0.0045
Performance Gain: Strong convergence with minimal overfitting at optimal step.
1from transformers import AutoProcessor, AutoModelForImageTextToText
2from PIL import Image
3import torch
45# Load model and processor6model_id ="naazimsnh02/dentalgemma-1.5-4b-it"7model = AutoModelForImageTextToText.from_pretrained(8 model_id,9 torch_dtype=torch.bfloat16,10 device_map="auto"11)12processor = AutoProcessor.from_pretrained(model_id)1314# Prepare input15image = Image.open("dental_xray.jpg")16messages =[17{18"role":"system",19"content":[{"type":"text","text":"You are an expert dental clinician and radiologist AI assistant."}]20},21{22"role":"user",23"content":[24{"type":"image","image": image},25{"type":"text","text":"Analyze this dental X-ray for any abnormalities."}26]27}28]2930# Generate response31inputs = processor.apply_chat_template(32 messages,33 add_generation_prompt=True,34 tokenize=True,35 return_dict=True,36 return_tensors="pt"37).to(model.device)3839outputs = model.generate(**inputs, max_new_tokens=300, do_sample=False)40generated_ids = outputs[0][inputs["input_ids"].shape[-1]:]41response = processor.decode(generated_ids, skip_special_tokens=True)42print(response)
Text-Only Inference (Clinical Cases)
python
1from transformers import AutoTokenizer, AutoModelForCausalLM
2import torch
34# Load model and tokenizer5model_id ="naazimsnh02/dentalgemma-1.5-4b-it"6tokenizer = AutoTokenizer.from_pretrained(model_id)7model = AutoModelForCausalLM.from_pretrained(8 model_id,9 torch_dtype=torch.bfloat16,10 device_map="auto"11)1213# Prepare clinical case14messages =[15{16"role":"system",17"content":"You are an expert dental clinician and radiologist AI assistant."18},19{20"role":"user",21"content":"A 35-year-old male presents with severe throbbing pain in the lower right molar region for 3 days. Clinical exam shows deep carious lesion on tooth #46 with tenderness to percussion. Periapical radiograph shows periapical radiolucency. Patient has no significant medical history. What is your assessment and management plan?"22}23]2425# Generate response26inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)27outputs = model.generate(inputs, max_new_tokens=512, do_sample=False)28response = tokenizer.decode(outputs[0], skip_special_tokens=True)29print(response)
🏗️ Model Architecture
DentalGemma inherits the architecture from MedGemma 1.5 4B IT:
Language Model: Gemma 3 architecture (4B parameters) — generates clinical text responses
Modality Fusion: Cross-attention mechanism for image-text integration
Context Length: 8,192 tokens
Vocabulary Size: 256,000 tokens
The LoRA fine-tuning adds trainable adapters (rank 64) to all linear layers. The final model merges these adapters back into the full-precision base weights for optimal inference performance.
📈 Evaluation
Qualitative Assessment
The model demonstrates strong performance on:
Clinical photo analysis with accurate cavity detection and severity assessment
Pathology classification from panoramic radiographs across 6 condition classes
Location-aware diagnosis with anatomical region identification (e.g., "right mandibular region")
Structured clinical reasoning following dental protocols with differential diagnoses
Treatment planning with appropriate antibiotic considerations and urgency assessment
Significant improvement from initial loss of ~1.32
Instruct Training (Stage 2):
Best Validation Loss: 0.0224 (at step 500)
~50% reduction in error compared to previous QLoRA runs (old best: 0.0435)
Training converged after ~2 epochs
Limitations
Not a diagnostic tool: This model is for research and educational purposes only. All AI-generated assessments must be validated by licensed dental professionals.
Training data bias: Performance may vary on X-ray types or clinical scenarios not well-represented in training data (e.g., bitewing radiographs, periapical films).
Hallucination risk: Like all LLMs, the model may occasionally generate plausible-sounding but incorrect information.
No real-time validation: The model cannot verify its outputs against current clinical guidelines or patient-specific contraindications.
Location accuracy: Dental region mapping from bounding boxes is approximate and should not be used for surgical planning without clinical verification.
⚠️ Ethical Considerations & Disclaimer
Intended Use
Research and education in dental AI applications
Proof-of-concept for domain adaptation of medical foundation models
Development tool for building dental diagnostic assistants
Out-of-Scope Use
Direct clinical diagnosis without professional oversight
Replacement for licensed dental professionals
Treatment decisions without clinical validation
Use in regulated medical devices without appropriate certification
Disclaimer
This model is NOT a substitute for professional dental diagnosis. All AI-generated assessments must be validated by licensed dental professionals before any clinical use. The model is provided "as-is" for research purposes only.
1@misc{medgemma2024,
2 title={MedGemma: Medical Foundation Models from Google Health},
3 author={Google Health AI},
4 year={2024},
5 publisher={HuggingFace},
6 howpublished={\url{https://huggingface.co/google/medgemma-1.5-4b-it}}
7}
Challenge Citation:
bibtex
1@misc{medgemma-impact-challenge,
2 author={Fereshteh Mahvar and Yun Liu and Daniel Golden and others},
3 title={The MedGemma Impact Challenge},
4 year={2026},
5 howpublished={\url{https://kaggle.com/competitions/med-gemma-impact-challenge}}
6}
