Comparison of two- and three-dimensional computed tomography-based deep-learning models with neurosurgeon assessments for predicting functional outcome at discharge in patients with spontaneous intracerebral hemorrhage: a single-center retrospective study

When a blood vessel bursts inside the brain, the resulting bleed is a medical emergency that strikes without warning. This event, known as a spontaneous intracerebral hemorrhage, is a leading cause of death and long-term disability worldwide. In the hours and days following such an event, doctors face a difficult task: predicting how well a patient will recover. To make these judgments, clinicians rely on two main sources of information. First, they examine non-contrast computed tomography scans, which are quick X-ray images that show the size and location of the bleeding. Second, they assess the patient's level of consciousness using a standard scoring system called the Glasgow Coma Scale, which measures how well a person can open their eyes, speak, and move in response to commands. While these tools are essential, human doctors must interpret them subjectively, weighing the visual evidence of the bleed against the patient's mental state. This process can vary from one specialist to another, and the complexity of the brain often hides subtle patterns that the human eye might miss.

A team of researchers at Longgang Central Hospital in Shenzhen set out to see if artificial intelligence could improve these predictions. They focused on a specific question: could a computer program, trained to look at brain scans and patient scores, predict a patient's functional outcome at the time of hospital discharge better than a human neurosurgeon could? The team did not just build one type of computer model; they built two. One model looked at the brain scans as a series of individual two-dimensional slices, much like flipping through the pages of a book. The other model looked at the entire brain as a single three-dimensional volume, treating the scan as a solid object. Both models were designed to combine the visual data from the scans with the clinical data from the patient's admission, specifically the Glasgow Coma Scale score. The goal was to see if these digital tools could spot the connection between the physical damage in the brain and the patient's likely recovery, and to determine if a simpler, slice-by-slice approach could perform just as well as the more complex, three-dimensional approach.

The researchers gathered data from 286 patients who had suffered a spontaneous brain bleed and were treated at their hospital between 2013 and 2021. For each patient, they collected the brain scan taken within 24 hours of the event and the Glasgow Coma Scale score recorded upon admission. They then divided the patients into groups to train and test their computer models. The team also recruited three neurosurgeons with different levels of experience to act as a human benchmark. These doctors were asked to predict the outcome for a selection of patients based on the same brain scans and clinical scores provided to the computers. To ensure a fair comparison, the doctors were shown only the specific slices of the scan containing the bleed, rather than the full series of images they would normally review in a hospital setting. This created a controlled environment where both the humans and the machines were working with the exact same information.

The results showed that the computer models were highly effective at predicting outcomes. The model that looked at the brain as a series of two-dimensional slices, when combined with the patient's Glasgow Coma Scale score, achieved an accuracy of 90.3 percent. This means that in more than nine out of ten cases, the model correctly predicted whether a patient would have a good or poor functional outcome at discharge. The three-dimensional model, which analyzed the entire brain volume along with more detailed clinical information, performed slightly differently, achieving an accuracy of 90.6 percent. Statistically, there was no significant difference between the two computer models; both were remarkably precise. However, both artificial intelligence systems outperformed the human neurosurgeons. Even the most experienced doctor in the study could not match the consistency and accuracy of the computer models. The study also found that adding the Glasgow Coma Scale score to the image analysis was crucial. A version of the two-dimensional model that looked only at the images without the clinical score was less accurate, dropping to 81.0 percent. This suggests that the computer needed the human context of the patient's consciousness level to make its best predictions.

One of the most surprising findings was that the simpler, two-dimensional model was just as good as the more complex, three-dimensional model. The three-dimensional approach required more computing power and more detailed data, yet it did not provide a better prediction than the slice-based method. This indicates that for this specific medical task, a less complex architecture might be preferable, especially when dealing with limited amounts of data. The researchers noted that the three-dimensional models are often more prone to errors when the dataset is small, whereas the two-dimensional approach remained robust. The study also highlighted that while the computers were excellent at finding patterns, they were not replacing the doctors. The models were designed to assist, not to take over. The human doctors in the study were working under a constrained review setting, looking at isolated images rather than the full clinical picture they would see in real life, which may have put them at a disadvantage compared to the computers that processed all the data at once.

The researchers were careful to point out the limits of their work. Because the study was conducted at a single hospital and looked back at past records, the results need to be tested in larger, forward-looking studies across multiple centers to ensure they hold true for all patients. The predictions made by the models were for the short-term outcome at the time of discharge, not for long-term recovery months or years later. Furthermore, the study confirmed that while the models are powerful, they are tools to support clinical decision-making, not to replace the judgment of a human specialist. The ultimate goal is to provide doctors with a reliable second opinion that can help them make more informed choices for patients facing a life-threatening brain bleed. By combining the visual evidence of the scan with the clinical reality of the patient's condition, these digital tools offer a new way to understand the severity of the injury and the likely path to recovery.