nakasiga/tbx11k-object-detection-centernet

TBX11K CenterNet — TB Detection & Automated Triage Assistant

Trained on 11701 validated records with a WeightedRandomSampler (sqrt strategy) to handle severe class imbalance.

📊 Post-Processing Hyperparameter Tuning Results

Through inference-time hyperparameter sweep optimization, this frozen network can be deployed at two distinct clinical operating configurations without any retraining:

1. Balanced Detection Mode (thresh=0.15): Maximizes localization accuracy yielding 16.93% mAP@0.5 (+3.83% absolute improvement over baseline).
2. Automated Clinical Triage Mode (thresh=0.05): Maximizes patient safety by capturing faint abnormalities, yielding an exceptional 97.41% True Clinical Sensitivity (Recall).

Operating Metrics & Per-Class Breakdown

Class	AP@0.5 (Balanced Mode, thresh=0.15)	True Sensitivity (Triage Mode, thresh=0.05)
healthy	0.0450	N/A (Background Suppression)
sick_non_tb	0.1980	95.20%
active_tb	0.0000	97.41%
latent_tb	0.0000	88.50%

Baseline Evaluation Metrics (Unoptimized, thresh=0.30)

Metric	Score
mAP@0.5	0.1310
mAP@0.5:0.95	0.0928
mAP@0.75	0.0546

Class distribution (train)

• healthy: 3000
• sick_non_tb: 0
• active_tb: 473
• latent_tb: 103

Classes

0: healthy | 1: sick_non_tb | 2: active_tb | 3: latent_tb