MICCAI 2022 MELA Challenge:

Mediastinal Lesion Analysis

Evaluation of Task: Detection

The evaluation of the detection performance is based on the Free-Response Receiver Operating Characteristic (FROC), which is an evaluation metric balancing both sensitivity and false positive. The FROC performance is reported as sensitivities at various false positive (FP) levels. In particular, the average sensitivity of  FP= 0.125, 0.25, 0.5, 1, 2, 4, 8 is adopted as the evaluation metric of the detection performance.

A detected proposal is regarded as a hit when the overlapping between the proposal and ground-truth bounding boxes of mediastinal lesions has an IoU>0.3. We choose the IoU>0.3, which is lower than IoU>0.5 as commonly used in 2D detection such as COCO. This is because 3D detection bounding boxes usually have lower IoU than those in 2D.

Please refer to the official evaluation code for the details of evaluation.

Requirements for Submission

You need to submit a .csv file including all your detected results for the test set. The prediction .csv should have eight columns: public_id (patient ID), coordX, coordY, coordZ (coordinates indicating the predicted bounding boxes of lesions), x_length, y_length, z_length (width, length, and depth of the predicted bounding boxes) and probability (detection confidence), e.g.:

The format of the required .csv submission.