在cvpr上少见的使用medical data的paper
Contributions
-
收集了新的很大的TB dataset：Tuberculosis X-ray (TBX11K) dataset，包括：
11200 X-ray Images
Image-level annotation + TB area annotation using bounding boxes
Image-level annotations include 4 classes: healthy, active TB, latent TB, & unhealthy but non-TB -
Reform existing object detectors to perform simultaneous image classification and TB area detection (SSD, RetinaNet, Faster-RCNN, FCOS)，并定义了classification 和 detection 的 metrics。 作为dataset的baselines
Methods
-
the classification branch learns to classify X-rays into 3 classes: healthy, sick but non-TB, and TB
evaluation metrics: accuracy, auc, sensitivity… -
the detection branch learns to detect TBs with 3 classes: active TB, latent TB
evaluation metrics: average precision of bounding box
Results
- 和其他datasets的对比，比其他大很多
- 作者对于每个baseline model都做了实验。从结果上看，Faster-RCNN 和 SSD 的表现比较突出。