DEEP LEARNING PIPELINE

Virtual Radiologist Assistant

AI-Powered Lower-Grade Glioma (LGG) Segmentation

Best Dice Score

0.8960 AttentionMetadataUNet + Scheduler

Best IoU Score

0.8135 AttentionMetadataUNet + Scheduler

Total MRI Slices

7,858 110 patients

Tumor Slices (Val)

535 of 1,572

Avg Tumor Coverage

3.0% tumor-positive slices

Pipeline Architecture

Phase	Model	Key Innovation
1	BasicUNet	Standard encoder-decoder
2	MetadataUNet	Clinical metadata fusion at bottleneck
3	AttentionMetadataUNet	Attention gates on skip connections
4	+ LR Scheduler	ReduceLROnPlateau for convergence

Training Configuration

Epochs	15
Batch Size	32
Learning Rate	0.001
Optimizer	Adam
Loss Function	DiceBCE (BCE + Dice)
Image Size	128 x 128

Final Model Scores

BasicUNet Dice

0.8941

MetadataUNet Dice

0.8949

AttentionMetadataUNet Dice

0.8718

Scheduler Dice

0.8960

REAL-TIME SEGMENTATION

Live MRI Inference

Upload a brain MRI slice and input patient metadata for real-time tumor segmentation.

Age at Diagnosis

Gender

Histological Type

Tumor Location

RNASeq Cluster

Methylation Cluster

Confidence Threshold

Ready. Select a sample or upload an MRI slice.

TRAINING INSIGHTS

Training Analytics

Validation Dice Score

BasicUNet MetadataUNet AttentionMetadataUNet Scheduler

BENCHMARK RESULTS

Model Comparison & Analysis

Final Validation Metrics

Model	Dice	IoU	Metadata	Attention	Scheduler
BasicUNet	0.894113	0.810121	No	No	No
MetadataUNet	0.894860	0.811078	Yes	No	No
AttentionMetadataUNet	0.871776	0.775612	Yes	Yes	No
AttentionMetadataUNet (Scheduler)	0.895969	0.813479	Yes	Yes	Yes

VISUAL RESULTS

Prediction Gallery

Side-by-side comparison of MRI slices, ground truth masks, model predictions, and overlays.

MRI Input

Ground Truth

Model Prediction

Tumor Overlay