Tbx11k Object Detection Yolov8

Per-class AP@0.5 Breakdown

Class Imbalance Optimization Strategy

• Grayscale-tuned heavier data augmentations (mosaic, spatial scaling, translation, brightness adjustments)
• Unannotated background control files maintained to minimize false positive bounding box proposals

Internal Network Class Configuration Map (0-Indexed)

• 0: Active TB
• 1: Latent TB (Note: 'Healthy' files carry no boxes and function natively as architectural background signal)

weights/epoch10.pt

pt · 198 MB · SHA-256 ef36073dffc6…2ce9 · Hugging Face

--- license: mit tags: - object-detection - medical - tuberculosis - yolov8 - chest-xray datasets: - tbx11k pipeline_tag: object-detection --- # TBX11K YOLOv8m — Tuberculosis Detection and Localization This model is a Ultralytics YOLOv8m detector fine-tuned on the TBX11K chest X-ray dataset for structural Tuberculosis localization and classification. ## Labeled Test Set Performance Evaluation Evaluated on an independent, labeled split extracted cleanly from the validation data stream to prevent zero-label evaluation matrix distortions. | Metric | Score | |--------|-------| | Test mAP@0.5 | 0.3551 | | Test mAP@0.5:0.95 | 0.1372 | ### Per-class AP@0.5 Breakdown | Class | AP@0.5 | |-------|--------| | active_tb | 0.5068 | | latent_tb | 0.2034 | ## Class Imbalance Optimization Strategy - Grayscale-tuned heavier data augmentations (mosaic, spatial scaling, translation, brightness adjustments) - Unannotated background control files maintained to minimize false positive bounding box proposals ## Internal Network Class Configuration Map (0-Indexed) - `0`: Active TB - `1`: Latent TB *(Note: 'Healthy' files carry no boxes and function natively as architectural background signal)*