Hemorrhages can be life-threatening emergencies that require immediate intervention. Rapid diagnosis and appropriate management are crucial for patient outcomes. In the case of severe hemorrhages, such as massive brain hemorrhages or hemoperitoneum, prompt medical attention can be critical in preventing further complications and saving lives. Hemorrhages have the potential to rapidly worsen if left untreated. For instance, in the case of a brain hemorrhage, the accumulating blood can exert pressure on the brain, leading to increased intracranial pressure and subsequent brain damage. In hemoperitoneum, uncontrolled bleeding can result in hypovolemic shock, organ failure, or even death. Timely diagnosis and intervention in the emergency department are vital to prevent such deteriorations

Deep learning-based automated detection or classification capability helps streamline the diagnostic process, saving valuable time for healthcare professionals in the emergency department. In this aspect, our challenges aim to the development of deep learning-based classification models for two targeted emergency conditions (i.e., brain hemorrhage and hemoperitoneum) on CT images.

Targeted emergency diseases

1) Brain hemorrhage, also known as intracranial hemorrhage, refers to bleeding that occurs within the brain tissue or the surrounding membranes. It is a critical medical emergency that requires urgent intervention and close monitoring.

2) Hemoperitoneum refers to the presence of blood in the peritoneal cavity, which is the space within the abdomen that contains various organs such as the liver, intestines, and stomach.

In this challenge, participants are tasked with performing image classification to distinguish between normal and abnormal images. The challenge involves two separate classification tasks using two different datasets.

1) Task 1 focuses on image classification for brain CT images, specifically distinguishing between images with and without brain hemorrhage.

2) Task 2 involves image classification for abdominopelvic CT images, specifically distinguishing between images with and without hemoperitoneum.

Participants will need to develop classification models that can accurately classify the images from both datasets, identifying the presence or absence of the target emergent diseases (brain hemorrhage and hemoperitoneum).

The dataset we deploy is a synthetic image created by a deep learning generative model trained with real medical image data. All data images are released under a CC-BY-NC-SA license. Synthetic data can be a valuable alternative when real data are scarce, or privacy concerns restrict their availability.

Abdomen Window-Setting (WL, WW): (35, 350)
Brain Window-Setting (WL, WW): (40, 80)

• Training and tuning data : All synthetic image data are weakly labeled as normal or abnormal, and is freely downloadable by all participants.
• Internal and external validation data : All validation data are real medical images and are only available on a cloud basis. They are provided to test the performance of the model, but is not freely downloadable.

The algorithms' performance will be evaluated using the following evaluation metrics.

Log Loss (Logarithmic Loss): Log Loss is a metric that measures the accuracy of a classification model's predicted probabilities. It quantifies how well the predicted probabilities align with the true class labels of the dataset. The basic idea behind Log Loss is to penalize models more heavily for making predictions that are confidently wrong.

Submissions are evaluated using the average of the sample weighted log losses.

Sample weights 1 for internal validation data (brain hemorrhage).
2 for external validation data (brain hemorrhage).

1 for internal validation data (hemoperitoneum).
2 for external validation data (hemoperitoneum).